Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: use ComfyUI's default settings for FLUX.1, Qwen-Image, Z-Image, Ideogram 4, Wan2.2 and HunyuanVideo-1.5 (#12516)

Daniel Hanfd2522bfc599
diff --git a/studio/backend/core/inference/diffusion.py b/studio/backend/core/inference/diffusion.py
index 3921b44d057..0def3f744e2 100644
--- a/studio/backend/core/inference/diffusion.py
+++ b/studio/backend/core/inference/diffusion.py
@@ -240,6 +240,15 @@ from .diffusion_precision import (
torchao_quantize_importable,
)
from .diffusion_te_prequant import te_prequant_pipe_kwargs
+from .diffusion_flow_shift import apply_comfy_flow_shift
+from .diffusion_text_length import (
+ IDEOGRAM4_COMFY_GUIDANCE,
+ IDEOGRAM4_COMFY_MU,
+ IDEOGRAM4_COMFY_STD,
+ ideogram4_comfy_guidance_schedule,
+ flux_t5_kwarg,
+ true_cfg_needs_empty_negative,
+)
from .diffusion_denoiser_prequant import (
DENOISER_COMPONENT,
PIPELINE_SEED_DECLINED,
@@ -6987,6 +6996,8 @@ class DiffusionBackend:
)
self._raise_if_load_cancelled(_load_token)
+ # Sample at ComfyUI's static sigma shift where the shipped scheduler differs (before from_pipe copies it).
+ apply_comfy_flow_shift(pipe, getattr(fam, "comfy_flow_shift", None), logger)
# Before the speed optims so their decode compile lands inside the non-finite check; `off` keeps fp32.
vae_fp16 = str(
speed_mode or ""
@@ -9405,7 +9416,20 @@ class DiffusionBackend:
if steps == 48 and abs(float(guidance) - 7.0) < 1e-6:
kwargs.pop(state.family.cfg_kwarg, None)
else:
- kwargs["guidance_schedule"] = None
+ # Otherwise ComfyUI's template: its "Default" schedule (mu 0.0, std 1.75; the card taper above
+ # keeps the pipeline's mu 0 / std 1.5) and, at the default guidance 7, its CFG override to 3
+ # over the last 30% of sampling, as a per-step schedule. Any other guidance stays constant.
+ if "mu" in call_params:
+ kwargs["mu"] = IDEOGRAM4_COMFY_MU
+ if "std" in call_params:
+ kwargs["std"] = IDEOGRAM4_COMFY_STD
+ if abs(float(guidance) - IDEOGRAM4_COMFY_GUIDANCE) < 1e-6:
+ kwargs.pop(state.family.cfg_kwarg, None)
+ kwargs["guidance_schedule"] = ideogram4_comfy_guidance_schedule(
+ steps, width, height
+ )
+ else:
+ kwargs["guidance_schedule"] = None
if state.family.name == LUMINA2_FAMILY_NAME and "cfg_trunc_ratio" in call_params:
# Lumina 2's card recipe truncates the CFG double-forward to the first quarter
# (cfg_trunc_ratio=0.25); the 1.0 default oversaturates.
@@ -9437,6 +9461,13 @@ class DiffusionBackend:
kwargs["height"] = ih
if negative_prompt and "negative_prompt" in call_params:
kwargs["negative_prompt"] = negative_prompt
+ elif "negative_prompt" in call_params and true_cfg_needs_empty_negative(
+ state.family.cfg_kwarg, guidance
+ ):
+ # Qwen-Image style pipelines run true CFG only when a negative is PRESENT, so a blank negative
+ # silently dropped CFG while the UI showed guidance 4. ComfyUI encodes an empty negative and
+ # applies CFG; do the same.
+ kwargs["negative_prompt"] = ""
if workflow == "controlnet" and control_pil is not None:
# CN pipeline takes the control map + scale; guidance start/end bound its step range. Every kwarg
# is signature-gated.
@@ -9635,6 +9666,11 @@ class DiffusionBackend:
chunk_kwargs["negative_prompt"] = [
chunk_kwargs["negative_prompt"]
] * len(chunk)
+ # FLUX.1 T5 length as ComfyUI pads it: 256 for prompts up to 256 tokens (diffusers pads every
+ # prompt to 512); longer prompts keep the 512 bucket. Per chunk, since a prompts list varies.
+ t5_len = flux_t5_kwarg(state.family.name, pipe, call_params, chunk_kwargs)
+ if t5_len is not None:
+ chunk_kwargs["max_sequence_length"] = t5_len
# A step cache keys residuals on the cond/uncond context, which a graph key
# cannot see. Per chunk because an AUTO decision is re-taken per generation.
if state.cuda_graphs:
diff --git a/studio/backend/core/inference/diffusion_families.py b/studio/backend/core/inference/diffusion_families.py
index 8889e1158e0..21bc614705e 100644
--- a/studio/backend/core/inference/diffusion_families.py
+++ b/studio/backend/core/inference/diffusion_families.py
@@ -113,6 +113,9 @@ class DiffusionFamily:
max_output_pixels: int = 2048 * 2048
# Accepted condition-image preprocessing resolutions (square side, by area); empty = no such control.
reference_resolutions: tuple[int, ...] = field(default_factory = tuple)
+ # Static sigma shift ComfyUI samples this family with; the loader rebuilds the scheduler at it. None = keep the
+ # shipped scheduler.
+ comfy_flow_shift: Optional[float] = None
# Activation-guard cost of one condition pixel relative to one output pixel.
condition_pixel_weight: float = 1.0
# Extra lowercased substrings (besides ``name``) that map a repo id here.
@@ -325,6 +328,7 @@ _FAMILIES: tuple[DiffusionFamily, ...] = (
# Qwen instruction editing: the 2511 checkpoint ships as QwenImageEditPlusPipeline. Specific aliases first so
# detect_family prefers this over "qwen-image".
name = "qwen-image-edit",
+ comfy_flow_shift = 3.1, # ComfyUI ModelSamplingAuraFlow 3.1 (Qwen-Image-Edit 2511 template)
pipeline_class = "QwenImageEditPlusPipeline",
transformer_class = "QwenImageTransformer2DModel",
base_repo = "Qwen/Qwen-Image-Edit-2511",
@@ -342,6 +346,7 @@ _FAMILIES: tuple[DiffusionFamily, ...] = (
),
DiffusionFamily(
name = "qwen-image",
+ comfy_flow_shift = 3.1, # ComfyUI ModelSamplingAuraFlow 3.1 (Qwen-Image templates)
pipeline_class = "QwenImagePipeline",
transformer_class = "QwenImageTransformer2DModel",
base_repo = "Qwen/Qwen-Image",
@@ -488,6 +493,7 @@ _FAMILIES: tuple[DiffusionFamily, ...] = (
),
DiffusionFamily(
name = "z-image",
+ comfy_flow_shift = 3.0, # ComfyUI shift 3 for Turbo and base (Turbo already ships 3.0)
pipeline_class = "ZImagePipeline",
transformer_class = "ZImageTransformer2DModel",
base_repo = "Tongyi-MAI/Z-Image-Turbo",
@@ -1132,29 +1138,37 @@ def prefer_ungated_mirror(
# Default (steps, guidance) per model for callers that cannot pass them. Matched by substring, most specific first;
# same values as the UI MODEL_DEFAULTS table, keep in sync.
_GENERATION_DEFAULTS: tuple[tuple[str, int, float], ...] = (
- ("z-image-turbo", 9, 0.0),
- # FLUX.1 Krea dev is a FLUX.1-dev finetune, NOT a Krea-2: 28 steps at guidance 4.5. Must precede the generic
- # "krea" key.
- ("flux.1-krea", 28, 4.5),
+ # Values follow ComfyUI's official templates for the same model (our baseline).
+ # Z-Image-Turbo: 8 steps (ComfyUI runs 8 sampler steps; diffusers runs one forward per step).
+ ("z-image-turbo", 8, 0.0),
+ # FLUX.1 Krea dev is a FLUX.1-dev finetune, NOT a Krea-2: 20 steps at the FLUX guidance default 3.5, as ComfyUI.
+ # Must precede the generic "krea" key.
+ ("flux.1-krea", 20, 3.5),
# Krea 2 Raw (undistilled): 52 steps / guidance 3.5. Must precede the generic "krea" key.
("krea-2-raw", 52, 3.5),
# Krea 2 Turbo (distilled): 8 steps, no CFG. "krea" then covers Turbo and other krea ids but Raw.
("krea", 8, 0.0),
("flux.1-schnell", 4, 0.0),
- ("kontext", 28, 2.5), # editing: before the generic flux.1
- ("flux.1", 28, 3.5),
- # The undistilled base variants need their model-card 50-step CFG recipe. Keep this before the generic distilled
- # key, which covers both 4B and 9B 4-step checkpoints.
- ("flux.2-klein-base", 50, 4.0),
+ ("kontext", 20, 2.5), # editing: before the generic flux.1
+ ("flux.1", 20, 3.5),
+ # The undistilled base variants run real CFG: 20 steps at CFG 5, as ComfyUI. Keep this before the generic
+ # distilled key, which covers both 4B and 9B 4-step checkpoints.
+ ("flux.2-klein-base", 20, 5.0),
("flux.2-klein", 4, 1.0),
- ("flux.2-dev", 28, 4.0), # full (non-distilled)
- # Qwen-Image-2.1: 40 steps, no guidance. Before the generic qwen-image key.
- ("qwen-image-2.1", 40, 1.0),
- ("qwen-image-21", 40, 1.0),
- ("qwen_image_21", 40, 1.0),
- ("qwenimage21", 40, 1.0),
+ ("flux.2-dev", 20, 4.0), # full (non-distilled)
+ # Qwen-Image-2.1: 25 steps, no guidance. Before the generic qwen-image key.
+ ("qwen-image-2.1", 25, 1.0),
+ ("qwen-image-21", 25, 1.0),
+ ("qwen_image_21", 25, 1.0),
+ ("qwenimage21", 25, 1.0),
+ # Qwen-Image-Edit 2511: 40 steps at CFG 4. Before the generic qwen-image key.
+ ("qwen-image-edit", 40, 4.0),
+ # Qwen-Image-2512: 50 steps at CFG 4 (its own ComfyUI template; the original Qwen-Image runs 20).
+ ("qwen-image-2512", 50, 4.0),
("qwen-image", 20, 4.0),
- ("z-image", 20, 4.0),
+ # Z-Image base: 25 steps at ComfyUI CFG 4. diffusers Z-Image computes pos + g * (pos - neg), so its g is
+ # ComfyUI's cfg - 1.
+ ("z-image", 25, 3.0),
# Lumina Image 2.0 card: 50 steps, guidance 4 (plus cfg_trunc_ratio 0.25, which the loader passes itself).
("lumina", 50, 4.0),
# HunyuanImage 2.1 card: 50 steps; guidance feeds distilled_guidance_scale, while real CFG runs inside the
@@ -1164,13 +1178,13 @@ _GENERATION_DEFAULTS: tuple[tuple[str, int, float], ...] = (
("hidream-i1-dev", 28, 0.0),
("hidream-i1-fast", 16, 0.0),
("hidream", 50, 5.0),
- # Ideogram 4 card: 48 steps, guidance 7 (its schedule tapers the last 3 steps; the loader keeps that taper at
- # these defaults).
- ("ideogram", 48, 7.0),
+ # Ideogram 4: ComfyUI's Default preset, 20 steps at guidance 7 (3 once sigma <= 0.3), logit-normal mu 0.0 / std 1.75.
+ # An explicit 48 steps at 7 still runs the card's tapered schedule.
+ ("ideogram", 20, 7.0),
# SDXL: Turbo distilled; base wants ~30 steps + CFG ~7. "sdxl-turbo" precedes "sdxl".
("sdxl-turbo", 3, 0.0),
- ("stable-diffusion-xl", 30, 7.0),
- ("sdxl", 30, 7.0),
+ ("stable-diffusion-xl", 25, 7.0),
+ ("sdxl", 25, 7.0),
)
# Unrecognised model: distilled few-step / no-CFG shape, matching the UI fallback.
_GENERATION_DEFAULT_FALLBACK = (9, 0.0)
diff --git a/studio/backend/core/inference/diffusion_flow_shift.py b/studio/backend/core/inference/diffusion_flow_shift.py
new file mode 100644
index 00000000000..e12d3cb0c2b
--- /dev/null
+++ b/studio/backend/core/inference/diffusion_flow_shift.py
@@ -0,0 +1,64 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Static sigma shift a family samples with in ComfyUI, applied to the loaded scheduler. No
+torch/diffusers imports: it only rebuilds ``pipe.scheduler`` from its own config.
+
+Several shipped schedulers shift sigmas differently from ComfyUI's defaults for the same model:
+Qwen-Image resolves a resolution-dependent exponential shift (about 2.0 at 1024 px) and stretches
+the tail to 0.02 where ComfyUI samples a constant 3.1, Z-Image base ships 6.0 against 3.0, Wan2.2
+ships 5.0 / 3.0 against 8 / 5 and HunyuanVideo-1.5 5.0 / 9.0 against 7. A flow-matching shift ``s``
+maps sigma to ``s * sigma / (1 + (s - 1) * sigma)`` in both, so setting the static shift (and
+dropping the dynamic shift and terminal stretch) reproduces ComfyUI's schedule.
+"""
+
+from __future__ import annotations
+
+from typing import Any, Optional
+
+
+def flow_shift_overrides(config: Any, shift: float) -> Optional[dict]:
+ """Scheduler config overrides that sample at static ``shift``, or None when the scheduler has
+ no flow shift Studio knows how to set."""
+ try:
+ keys = set(config.keys())
+ except Exception: # noqa: BLE001 - an exotic config just keeps its scheduler
+ return None
+ if "flow_shift" in keys and config.get("use_flow_sigmas", True):
+ # UniPC / DPM-Solver in flow mode (Wan).
+ return {"flow_shift": float(shift)}
+ if "shift" in keys and "use_dynamic_shifting" in keys:
+ # FlowMatchEulerDiscrete: static shift, no resolution-dependent mu, no terminal stretch.
+ overrides = {"shift": float(shift), "use_dynamic_shifting": False}
+ if "shift_terminal" in keys:
+ overrides["shift_terminal"] = None
+ return overrides
+ return None
+
+
+def apply_comfy_flow_shift(
+ pipe: Any,
+ shift: Optional[float],
+ logger: Any = None,
+) -> bool:
+ """Rebuild ``pipe.scheduler`` at ComfyUI's static ``shift``. True when it changed."""
+ if shift is None:
+ return False
+ scheduler = getattr(pipe, "scheduler", None)
+ config = getattr(scheduler, "config", None)
+ if scheduler is None or config is None:
+ return False
+ overrides = flow_shift_overrides(config, shift)
+ if not overrides:
+ return False
+ if all(config.get(k) == v for k, v in overrides.items()):
+ return False
+ try:
+ pipe.scheduler = type(scheduler).from_config(config, **overrides)
+ except Exception as exc: # noqa: BLE001 - keep the shipped schedule rather than fail the load
+ if logger is not None:
+ logger.warning("flow shift %s not applied: %s", shift, exc)
+ return False
+ if logger is not None:
+ logger.info("scheduler flow shift set to %s (ComfyUI default)", shift)
+ return True
diff --git a/studio/backend/core/inference/diffusion_text_length.py b/studio/backend/core/inference/diffusion_text_length.py
new file mode 100644
index 00000000000..0ece9c90116
--- /dev/null
+++ b/studio/backend/core/inference/diffusion_text_length.py
@@ -0,0 +1,147 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Text-conditioning defaults that follow ComfyUI. No torch/diffusers imports.
+
+FLUX.1 T5 sequence length:
+
+diffusers pads (and truncates) the FLUX.1 T5 prompt to ``max_sequence_length=512`` and runs T5
+without an attention mask, so every pad token takes part in the encoder's self-attention. ComfyUI,
+our baseline, pads the same T5 prompt only up to 256 tokens and otherwise uses its real length
+(again with no mask). Both the embeddings and the text tokens the DiT attends over therefore differ,
+and 512 costs 256 extra joint-attention tokens per step for a short prompt.
+
+``flux_t5_sequence_length`` reproduces the ComfyUI length (256) for every prompt of at most 256 T5
+tokens (EOS included). Longer prompts pad to 512 as before (bucketed, see the function), and a prompt
+past 512 tokens truncates exactly as before.
+"""
+
+from __future__ import annotations
+
+import math
+from typing import Any, Iterable, Optional
+
+# ComfyUI's Ideogram 4 template ("Default" preset): 20 steps, logit-normal mean 0.0 and spread 1.75 before the
+# resolution term (both implementations add it the same way), and guidance 7 overridden to 3 over the last 30% of
+# sampling (CFG override, start 0.7 / end 1.0). Ideogram 4 samples a plain flow model (shift 1), so that range is
+# sigma <= 0.3.
+IDEOGRAM4_COMFY_STEPS = 20
+IDEOGRAM4_COMFY_MU = 0.0
+IDEOGRAM4_COMFY_STD = 1.75
+IDEOGRAM4_COMFY_GUIDANCE = 7.0
+IDEOGRAM4_COMFY_TAIL_GUIDANCE = 3.0
+IDEOGRAM4_COMFY_TAIL_SIGMA = 0.3
+_IDEOGRAM4_LOGSNR_MIN = -15.0
+_IDEOGRAM4_LOGSNR_MAX = 18.0
+
+
+def ideogram4_sigmas(steps: int, width: int, height: int, mu: float, std: float) -> list[float]:
+ """The ``steps`` sigmas the Ideogram 4 logit-normal schedule visits (terminal 0 excluded), highest first."""
+ from statistics import NormalDist
+
+ mean = float(mu) + 0.5 * math.log((int(width) * int(height)) / (512 * 512))
+ t_min = 1.0 / (1.0 + math.exp(0.5 * _IDEOGRAM4_LOGSNR_MAX))
+ t_max = 1.0 / (1.0 + math.exp(0.5 * _IDEOGRAM4_LOGSNR_MIN))
+ out = []
+ for i in range(int(steps)):
+ u = 1.0 - i / int(steps) # the schedule is built on linspace(0, 1) and flipped
+ if u >= 1.0:
+ t = 0.0
+ else:
+ y = mean + float(std) * NormalDist().inv_cdf(u)
+ t = 1.0 - 1.0 / (1.0 + math.exp(-y))
+ out.append(1.0 - min(max(t, t_min), t_max))
+ return out
+
+
+def ideogram4_comfy_guidance_schedule(steps: int, width: int, height: int) -> list[float]:
+ """Per-step guidance ComfyUI's template applies: 7, then 3 on every step whose sigma is <= 0.3."""
+ return [
+ IDEOGRAM4_COMFY_TAIL_GUIDANCE
+ if sigma <= IDEOGRAM4_COMFY_TAIL_SIGMA
+ else IDEOGRAM4_COMFY_GUIDANCE
+ for sigma in ideogram4_sigmas(steps, width, height, IDEOGRAM4_COMFY_MU, IDEOGRAM4_COMFY_STD)
+ ]
+
+
+# FLUX.1 families on a FluxPipeline-style ``max_sequence_length`` whose ComfyUI tokenizer is the FLUX T5 one.
+FLUX_T5_FAMILIES = frozenset({"flux.1", "flux.1-kontext"})
+FLUX_T5_MIN_TOKENS = 256
+FLUX_T5_MAX_TOKENS = 512
+
+
+def _texts(value: Any) -> list[str]:
+ if value is None:
+ return []
+ if isinstance(value, str):
+ return [value]
+ if isinstance(value, (list, tuple)):
+ return [v for v in value if isinstance(v, str)]
+ return []
+
+
+def t5_token_count(tokenizer: Any, text: str) -> int:
+ """T5 tokens for ``text`` including the EOS the tokenizer appends."""
+ ids = tokenizer(text, add_special_tokens = True)["input_ids"]
+ return len(ids)
+
+
+def flux_t5_sequence_length(
+ tokenizer: Any,
+ prompts: Iterable[Any],
+ *,
+ floor: int = FLUX_T5_MIN_TOKENS,
+ cap: int = FLUX_T5_MAX_TOKENS,
+) -> Optional[int]:
+ """``max_sequence_length`` for one FLUX.1 call, or None when it cannot be measured (keep the
+ pipeline default). ``prompts`` holds each str / list-of-str the call encodes with T5."""
+ if tokenizer is None or not callable(tokenizer):
+ return None
+ texts: list[str] = []
+ for value in prompts:
+ texts.extend(_texts(value))
+ if not texts:
+ return None
+ try:
+ longest = max(t5_token_count(tokenizer, t) for t in texts)
+ except Exception: # noqa: BLE001 - an odd tokenizer only keeps the pipeline default
+ return None
+ # Bucketed, a deliberate deviation from ComfyUI's exact length: every distinct text length is a new denoiser shape,
+ # and on the compiled default path the first one past 256 measured a 14 s recompile, each further one a CUDA graph
+ # recapture (0.4 to 1 s) until the 4-graph cap made later shapes run eager. So a prompt past 256 tokens pads to the
+ # cap (exactly the old behaviour); T5 runs unmasked, so that padding shifts the embeddings of those prompts only.
+ return int(floor) if longest <= int(floor) else int(cap)
+
+
+def flux_t5_kwarg(
+ family_name: str, pipe: Any, call_params: Any, chunk_kwargs: dict
+) -> Optional[int]:
+ """The ``max_sequence_length`` to pass for this chunk, or None to leave the kwarg alone.
+
+ Only FLUX.1 families, only when the pipeline accepts the kwarg, and never over a value the
+ caller already set. The negative prompt counts only when true CFG actually encodes it."""
+ if family_name not in FLUX_T5_FAMILIES:
+ return None
+ if "max_sequence_length" not in call_params or "max_sequence_length" in chunk_kwargs:
+ return None
+ prompts: list[Any] = [chunk_kwargs.get("prompt_2") or chunk_kwargs.get("prompt")]
+ negative = chunk_kwargs.get("negative_prompt_2") or chunk_kwargs.get("negative_prompt")
+ try:
+ true_cfg = float(chunk_kwargs.get("true_cfg_scale", 1.0) or 1.0)
+ except (TypeError, ValueError):
+ true_cfg = 1.0
+ if negative and true_cfg > 1.0:
+ prompts.append(negative)
+ return flux_t5_sequence_length(getattr(pipe, "tokenizer_2", None), prompts)
+
+
+def true_cfg_needs_empty_negative(cfg_kwarg: str, guidance: Any) -> bool:
+ """True when a ``true_cfg_scale`` pipeline (Qwen-Image family) would skip CFG for want of a
+ negative prompt. diffusers enables true CFG only when ``true_cfg_scale > 1`` AND a negative is
+ given; ComfyUI always encodes the (possibly empty) negative and applies CFG above 1."""
+ if cfg_kwarg != "true_cfg_scale":
+ return False
+ try:
+ return float(guidance) > 1.0
+ except (TypeError, ValueError):
+ return False
diff --git a/studio/backend/core/inference/sd_cpp_backend.py b/studio/backend/core/inference/sd_cpp_backend.py
index e94b068d6ad..7cf0df65f55 100644
--- a/studio/backend/core/inference/sd_cpp_backend.py
+++ b/studio/backend/core/inference/sd_cpp_backend.py
@@ -2175,6 +2175,9 @@ def _map_guidance(
``--cfg-scale``. A distilled 0/1 means CFG off (sd-cli's 1.0); a value > 1 is real CFG."""
if fam.name in ("flux.1", "flux.2-klein", "flux.2-dev"):
return None, (float(guidance) if guidance is not None else None)
+ if fam.name == "z-image":
+ # diffusers Z-Image computes pos + g * (pos - neg), so its g is standard CFG minus 1 (sd.cpp's cfg 4 == g 3).
+ return (float(guidance) + 1.0 if guidance is not None and guidance > 0.0 else 1.0), None
cfg = float(guidance) if (guidance is not None and guidance > 1.0) else 1.0
return cfg, None
diff --git a/studio/backend/core/inference/video.py b/studio/backend/core/inference/video.py
index 44b1ec385a2..388f4977765 100644
--- a/studio/backend/core/inference/video.py
+++ b/studio/backend/core/inference/video.py
@@ -55,6 +55,7 @@ from .diffusion_attention import (
sdpa_math_only,
select_attention_backend,
)
+from .diffusion_flow_shift import apply_comfy_flow_shift
from .diffusion_cache import (
FBCACHE_MIN_STEPS,
TC_AUTO,
@@ -6532,6 +6533,8 @@ class VideoBackend:
}
)
+ # Sample at ComfyUI's static sigma shift where the shipped scheduler differs.
+ apply_comfy_flow_shift(pipe, getattr(fam, "comfy_flow_shift", None), logger)
from . import diffusion_prompt_cache
diffusion_prompt_cache.install(
@@ -7499,6 +7502,8 @@ class VideoBackend:
),
}
)
+ # Sample at ComfyUI's static sigma shift where the shipped scheduler differs.
+ apply_comfy_flow_shift(pipe, getattr(fam, "comfy_flow_shift", None), logger)
from . import diffusion_prompt_cache
diffusion_prompt_cache.install(
diff --git a/studio/backend/core/inference/video_families.py b/studio/backend/core/inference/video_families.py
index 0739f3c2e61..a3d5bad2f7d 100644
--- a/studio/backend/core/inference/video_families.py
+++ b/studio/backend/core/inference/video_families.py
@@ -130,6 +130,9 @@ class VideoFamily:
modular_workflow: Optional[str] = None
# Released video and audio sigma shifts, when configurable.
default_flow_shift: Optional[float] = None
+ # Static sigma shift ComfyUI samples this family with by default; the loader rebuilds the scheduler at it. None =
+ # keep the shipped scheduler. Distinct from default_flow_shift, which also exposes a user control.
+ comfy_flow_shift: Optional[float] = None
default_audio_flow_shift: Optional[float] = None
# First/last-frame conditioning: the request may carry keyframe images.
supports_keyframes: bool = False
@@ -263,9 +266,10 @@ _FAMILIES: tuple[VideoFamily, ...] = (
prequant_variant_repos = (("lightricks/ltx-2.3", "fp8", "unsloth/LTX-2.3-FP8"),),
),
# Wan2.2-TI2V-5B (diffusers >= 0.35, verified on 0.39): ~5B single-stream DiT (UMT5 encoder), no audio. Its VAE's
- # temporal compression 4 gives valid frame counts 4k+1. Defaults 50 steps / CFG 5.
+ # temporal compression 4 gives valid frame counts 4k+1. Defaults 20 steps / CFG 5 (ComfyUI's template).
VideoFamily(
name = "wan2.2-ti2v-5b",
+ comfy_flow_shift = 8.0, # ComfyUI ModelSamplingSD3 8 (TI2V-5B template)
pipeline_class = "WanPipeline",
transformer_class = "WanTransformer3DModel",
base_repo = "Wan-AI/Wan2.2-TI2V-5B-Diffusers",
@@ -275,7 +279,7 @@ _FAMILIES: tuple[VideoFamily, ...] = (
# "wan2.2-5b"/"wan-ti2v" are the picker/GGUF short ids; "wan2.2-ti2v" catches the repo stem
aliases = ("wan2.2-5b", "wan-ti2v", "wan2.2-ti2v", "wan-ti2v-5b"),
has_audio = False,
- default_steps = 50,
+ default_steps = 20,
default_guidance = 5.0,
default_num_frames = 121,
default_fps = 24,
@@ -298,6 +302,7 @@ _FAMILIES: tuple[VideoFamily, ...] = (
# transformer_2, so cfg2_kwarg is threaded only here.
VideoFamily(
name = "wan2.2-t2v-a14b",
+ comfy_flow_shift = 5.0, # ComfyUI ModelSamplingSD3 5 (T2V-A14B template)
pipeline_class = "WanPipeline",
transformer_class = "WanTransformer3DModel",
base_repo = "Wan-AI/Wan2.2-T2V-A14B-Diffusers",
@@ -315,8 +320,9 @@ _FAMILIES: tuple[VideoFamily, ...] = (
transformer2_class = "WanTransformer3DModel",
is_moe = True,
cfg2_kwarg = "guidance_scale_2",
- default_steps = 50,
- default_guidance = 5.0,
+ # ComfyUI's A14B template without the Lightning LoRA: 20 steps at CFG 3.5 on both experts.
+ default_steps = 20,
+ default_guidance = 3.5,
# 81 frames at 16 fps ~5s (81 = 4*20 + 1), the A14B card's default clip.
default_num_frames = 81,
default_fps = 16, # A14B runs at 16 fps (vs TI2V-5B's 24)
@@ -335,6 +341,7 @@ _FAMILIES: tuple[VideoFamily, ...] = (
# so only the community repacks load.
VideoFamily(
name = "hunyuanvideo-1.5",
+ comfy_flow_shift = 7.0, # ComfyUI model default and template shift 7
pipeline_class = "HunyuanVideo15Pipeline",
transformer_class = "HunyuanVideo15Transformer3DModel",
base_repo = "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v",
@@ -345,7 +352,7 @@ _FAMILIES: tuple[VideoFamily, ...] = (
aliases = ("hunyuanvideo-1-5", "hunyuanvideo1.5", "hunyuanvideo1-5", "hv15"),
has_audio = False,
guidance_via_guider = True,
- default_steps = 50,
+ default_steps = 20,
default_guidance = 6.0,
default_num_frames = 121,
default_fps = 24,
@@ -364,6 +371,7 @@ _FAMILIES: tuple[VideoFamily, ...] = (
# own family so a 720p load defaults to 720p sizes; the full-path alias outranks the generic token.
VideoFamily(
name = "hunyuanvideo-1.5-720p",
+ comfy_flow_shift = 7.0, # ComfyUI model default and template shift 7
pipeline_class = "HunyuanVideo15Pipeline",
transformer_class = "HunyuanVideo15Transformer3DModel",
base_repo = "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v",
@@ -374,7 +382,7 @@ _FAMILIES: tuple[VideoFamily, ...] = (
aliases = ("hunyuanvideo-1.5-diffusers-720p_t2v", "hv15-720p"),
has_audio = False,
guidance_via_guider = True,
- default_steps = 50,
+ default_steps = 20,
default_guidance = 6.0,
default_num_frames = 121,
default_fps = 24,
@@ -809,10 +817,14 @@ def validate_video_reference_conditioning(
_VIDEO_GENERATION_DEFAULTS: tuple[tuple[str, int, float], ...] = (
("distilled", 8, 1.0),
("ltx", 40, 4.0),
- # Wan2.2 pipelines default to 50 steps / CFG 5.0; both TI2V-5B and A14B share these.
- ("wan", 50, 5.0),
- # HunyuanVideo-1.5: 50 steps with the guider's shipped CFG 6.0.
- ("hunyuanvideo", 50, 6.0),
+ # ComfyUI's templates for the same models (our baseline). Wan2.2-T2V-A14B: 20 steps at CFG 3.5 (Lightning LoRA
+ # off); before the generic Wan key.
+ ("a14b", 20, 3.5),
+ ("wan2.2-14b", 20, 3.5),
+ # Wan2.2-TI2V-5B: 20 steps at CFG 5.
+ ("wan", 20, 5.0),
+ # HunyuanVideo-1.5: 20 steps with the guider's shipped CFG 6.0.
+ ("hunyuanvideo", 20, 6.0),
)
diff --git a/studio/backend/tests/test_diffusion_backend.py b/studio/backend/tests/test_diffusion_backend.py
index 2d41b06e38e..ba78ff05810 100644
--- a/studio/backend/tests/test_diffusion_backend.py
+++ b/studio/backend/tests/test_diffusion_backend.py
@@ -14420,3 +14420,148 @@ def test_diffusion_status_response_keeps_the_gguf_a_swap_replaced():
]
assert dumped["replaced"] == "gguf:m-Q4_K_M.gguf"
assert dumped["artifact"] == "prequant:o/r/f.safetensors"
+
+
+# FLUX.1 T5 length (ComfyUI parity): real prompt length floored at 256, capped at 512.
+class _T5WordTokenizer:
+ """One id per whitespace word plus EOS, like T5TokenizerFast on plain words."""
+
+ def __call__(
+ self,
+ text,
+ add_special_tokens = True,
+ **_,
+ ):
+ return {"input_ids": [5] * len(text.split()) + ([1] if add_special_tokens else [])}
+
+
+class _FluxFakePipe(_FakePipe):
+ def __init__(self):
+ super().__init__()
+ self.tokenizer_2 = _T5WordTokenizer()
+
+ def __call__(
+ self,
+ *,
+ prompt = None,
+ max_sequence_length = 512,
+ **kwargs,
+ ):
+ return super().__call__(prompt = prompt, max_sequence_length = max_sequence_length, **kwargs)
+
+
+def test_generate_passes_flux1_t5_length_like_comfy(fake_runtime, tmp_path, monkeypatch):
+ diffusers = sys.modules["diffusers"]
+ monkeypatch.setattr(diffusers, "FluxPipeline", _FakePipeline, raising = False)
+ monkeypatch.setattr(diffusers, "FluxTransformer2DModel", _FakeTransformer, raising = False)
+ _no_cache(monkeypatch)
+ backend = _loaded_backend(tmp_path, family_override = "flux.1")
+ pipe = _FluxFakePipe()
+ object.__setattr__(backend._state, "pipe", pipe)
+ backend.generate(prompt = "a sloth on a branch", steps = 4, guidance = 0.0)
+ assert pipe.last_kwargs["max_sequence_length"] == 256
+ backend.generate(prompt = " ".join(["w"] * 320), steps = 4, guidance = 0.0)
+ assert pipe.last_kwargs["max_sequence_length"] == 512 # past 256 tokens: the 512 bucket
+
+
+def test_generate_leaves_t5_length_alone_off_flux1(fake_runtime, tmp_path):
+ backend = _loaded_backend(tmp_path, family_override = "z-image")
+ pipe = _FluxFakePipe()
+ object.__setattr__(backend._state, "pipe", pipe)
+ backend.generate(prompt = "a sloth", steps = 4, guidance = 0.0)
+ assert pipe.last_kwargs["max_sequence_length"] == 512 # pipeline default, nothing passed
+
+
+def test_qwen_true_cfg_gets_an_empty_negative_like_comfy(fake_runtime, tmp_path, monkeypatch):
+ """diffusers runs Qwen-Image true CFG only when a negative is present; ComfyUI always encodes the
+ empty negative and applies CFG. A blank negative must not silently turn CFG off."""
+ diffusers = sys.modules["diffusers"]
+ monkeypatch.setattr(diffusers, "QwenImagePipeline", _FakePipeline, raising = False)
+ monkeypatch.setattr(diffusers, "QwenImageTransformer2DModel", _FakeTransformer, raising = False)
+ _no_cache(monkeypatch)
+ backend = _loaded_backend(tmp_path, family_override = "qwen-image")
+ backend.generate(prompt = "a sloth", steps = 4, guidance = 4.0)
+ call = backend._state.pipe.last_kwargs
+ assert call["true_cfg_scale"] == 4.0 and call["negative_prompt"] == ""
+ # An explicit negative is kept; guidance <= 1 never asks for CFG.
+ backend.generate(prompt = "a sloth", negative_prompt = "blurry", steps = 4, guidance = 4.0)
+ assert backend._state.pipe.last_kwargs["negative_prompt"] == "blurry"
+ backend.generate(prompt = "a sloth", steps = 4, guidance = 1.0)
+ assert backend._state.pipe.last_kwargs["negative_prompt"] is None
+
+
+def test_guidance_scale_families_get_no_injected_negative(fake_runtime, tmp_path):
+ backend = _loaded_backend(tmp_path) # z-image: guidance_scale, its own CFG handling
+ backend.generate(prompt = "a sloth", steps = 4, guidance = 4.0)
+ assert backend._state.pipe.last_kwargs["negative_prompt"] is None
+
+
+class _IdeogramScheduleFakePipe(_FakePipe):
+ def __call__(
+ self,
+ *,
+ prompt = None,
+ mu = 0.0,
+ std = 1.5,
+ guidance_schedule = "card",
+ **kwargs,
+ ):
+ return super().__call__(
+ prompt = prompt, mu = mu, std = std, guidance_schedule = guidance_schedule, **kwargs
+ )
+
+
+def test_generate_ideogram_defaults_follow_comfy_template(fake_runtime, tmp_path):
+ """ComfyUI's Ideogram 4 template: 20 steps, mu 0.0 / std 1.75, guidance 7 overridden to 3 where sigma <= 0.3
+ (the last 3 of 20 steps at 1024^2). Another guidance stays constant; an explicit 48 / 7 keeps the card taper."""
+ backend = DiffusionBackend()
+ _load_ideogram(backend, tmp_path)
+ pipe = _IdeogramScheduleFakePipe()
+ object.__setattr__(backend._state, "pipe", pipe)
+ backend.generate(prompt = "a sloth", width = 1024, height = 1024, steps = 20, guidance = 7.0)
+ call = pipe.last_kwargs
+ assert call["guidance_scale"] is None
+ assert call["guidance_schedule"] == [7.0] * 17 + [3.0] * 3
+ assert (call["mu"], call["std"]) == (0.0, 1.75)
+ backend.generate(prompt = "a sloth", width = 512, height = 512, steps = 20, guidance = 7.0)
+ assert pipe.last_kwargs["guidance_schedule"] == [7.0] * 14 + [3.0] * 6
+ backend.generate(prompt = "a sloth", width = 1024, height = 1024, steps = 20, guidance = 5.0)
+ call = pipe.last_kwargs
+ assert call["guidance_scale"] == 5.0 and call["guidance_schedule"] is None
+ assert (call["mu"], call["std"]) == (0.0, 1.75)
+ backend.generate(prompt = "a sloth", steps = 48, guidance = 7.0)
+ call = pipe.last_kwargs
+ assert call["guidance_schedule"] == "card" and (call["mu"], call["std"]) == (0.0, 1.5)
+
+
+class _ShiftSchedulerConfig(dict):
+ pass
+
+
+class _ShiftFakeScheduler:
+ def __init__(self, **config):
+ self.config = _ShiftSchedulerConfig(config)
+
+ @classmethod
+ def from_config(cls, config, **overrides):
+ return cls(**{**config, **overrides})
+
+
+class _QwenShiftFakePipeline(_FakePipeline):
+ @classmethod
+ def from_pretrained(cls, base, **kwargs):
+ pipe = super().from_pretrained(base, **kwargs)
+ pipe.scheduler = _ShiftFakeScheduler(
+ shift = 1.0, use_dynamic_shifting = True, shift_terminal = 0.02
+ )
+ return pipe
+
+
+def test_qwen_load_samples_at_comfy_static_shift(fake_runtime, tmp_path, monkeypatch):
+ diffusers = sys.modules["diffusers"]
+ monkeypatch.setattr(diffusers, "QwenImagePipeline", _QwenShiftFakePipeline, raising = False)
+ monkeypatch.setattr(diffusers, "QwenImageTransformer2DModel", _FakeTransformer, raising = False)
+ _no_cache(monkeypatch)
+ backend = _loaded_backend(tmp_path, family_override = "qwen-image")
+ cfg = backend._state.pipe.scheduler.config
+ assert (cfg["shift"], cfg["use_dynamic_shifting"], cfg["shift_terminal"]) == (3.1, False, None)
diff --git a/studio/backend/tests/test_diffusion_flow_shift.py b/studio/backend/tests/test_diffusion_flow_shift.py
new file mode 100644
index 00000000000..a50392db789
--- /dev/null
+++ b/studio/backend/tests/test_diffusion_flow_shift.py
@@ -0,0 +1,73 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Families sample at ComfyUI's static sigma shift: the loader rebuilds the shipped scheduler."""
+
+from __future__ import annotations
+
+import types
+
+from core.inference.diffusion_families import detect_family
+from core.inference.diffusion_flow_shift import apply_comfy_flow_shift, flow_shift_overrides
+from core.inference.video_families import detect_video_family
+
+
+class _Config(dict):
+ pass
+
+
+class _FakeScheduler:
+ def __init__(self, **config):
+ self.config = _Config(config)
+
+ @classmethod
+ def from_config(cls, config, **overrides):
+ return cls(**{**config, **overrides})
+
+
+# Shipped configs (the fields that matter) of the Qwen-Image, Wan2.2-TI2V-5B and HunyuanVideo-1.5 720p repos.
+_QWEN = dict(
+ shift = 1.0, use_dynamic_shifting = True, shift_terminal = 0.02, time_shift_type = "exponential"
+)
+_WAN = dict(flow_shift = 5.0, use_flow_sigmas = True, solver_order = 2)
+_HV15 = dict(shift = 9.0, use_dynamic_shifting = False, shift_terminal = None)
+
+
+def test_flow_match_gets_a_static_shift_without_terminal_stretch():
+ pipe = types.SimpleNamespace(scheduler = _FakeScheduler(**_QWEN))
+ assert apply_comfy_flow_shift(pipe, 3.1) is True
+ cfg = pipe.scheduler.config
+ assert (cfg["shift"], cfg["use_dynamic_shifting"], cfg["shift_terminal"]) == (3.1, False, None)
+ assert cfg["time_shift_type"] == "exponential" # untouched keys survive
+ assert apply_comfy_flow_shift(pipe, 3.1) is False # idempotent
+
+
+def test_unipc_flow_shift_and_static_euler():
+ wan = types.SimpleNamespace(scheduler = _FakeScheduler(**_WAN))
+ assert apply_comfy_flow_shift(wan, 8.0) and wan.scheduler.config["flow_shift"] == 8.0
+ hv = types.SimpleNamespace(scheduler = _FakeScheduler(**_HV15))
+ assert apply_comfy_flow_shift(hv, 7.0) and hv.scheduler.config["shift"] == 7.0
+
+
+def test_no_shift_or_unknown_scheduler_is_a_no_op():
+ pipe = types.SimpleNamespace(scheduler = _FakeScheduler(**_QWEN))
+ assert apply_comfy_flow_shift(pipe, None) is False
+ other = types.SimpleNamespace(scheduler = _FakeScheduler(beta_schedule = "scaled_linear"))
+ assert apply_comfy_flow_shift(other, 3.0) is False
+ assert flow_shift_overrides(_Config(beta_schedule = "x"), 3.0) is None
+ assert apply_comfy_flow_shift(types.SimpleNamespace(), 3.0) is False
+
+
+def test_family_shifts_match_comfy_defaults():
+ assert detect_family("Qwen/Qwen-Image").comfy_flow_shift == 3.1
+ assert detect_family("Qwen/Qwen-Image-Edit-2511").comfy_flow_shift == 3.1
+ assert detect_family("Tongyi-MAI/Z-Image").comfy_flow_shift == 3.0
+ # Families whose shipped schedule already matches keep it.
+ assert detect_family("black-forest-labs/FLUX.1-schnell").comfy_flow_shift is None
+ assert detect_video_family("Wan-AI/Wan2.2-TI2V-5B-Diffusers").comfy_flow_shift == 8.0
+ assert detect_video_family("Wan-AI/Wan2.2-T2V-A14B-Diffusers").comfy_flow_shift == 5.0
+ for repo in (
+ "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v",
+ "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v",
+ ):
+ assert detect_video_family(repo).comfy_flow_shift == 7.0
diff --git a/studio/backend/tests/test_diffusion_flux_t5_length.py b/studio/backend/tests/test_diffusion_flux_t5_length.py
new file mode 100644
index 00000000000..993b7ea29d0
--- /dev/null
+++ b/studio/backend/tests/test_diffusion_flux_t5_length.py
@@ -0,0 +1,84 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""FLUX.1 T5 sequence length follows ComfyUI: the prompt's real T5 length (EOS included), floored
+at 256 and capped at the pipeline's 512, instead of diffusers' fixed 512."""
+
+from __future__ import annotations
+
+import types
+
+import pytest
+
+from core.inference.diffusion_text_length import (
+ flux_t5_kwarg,
+ flux_t5_sequence_length,
+)
+
+
+class _WordTokenizer:
+ """One id per whitespace word plus EOS, like T5TokenizerFast on plain words."""
+
+ def __call__(
+ self,
+ text,
+ add_special_tokens = True,
+ **_,
+ ):
+ ids = [5] * len(text.split())
+ return {"input_ids": ids + ([1] if add_special_tokens else [])}
+
+
+def _words(n):
+ return " ".join(["w"] * n)
+
+
+ "words, expected",
+ [(3, 256), (255, 256), (256, 512), (300, 512), (511, 512), (700, 512)],
+)
+def test_length_is_256_up_to_256_tokens_else_the_512_bucket(words, expected):
+ assert flux_t5_sequence_length(_WordTokenizer(), [_words(words)]) == expected
+
+
+def test_longest_prompt_of_a_list_sets_the_length():
+ assert flux_t5_sequence_length(_WordTokenizer(), [["a", _words(200)]]) == 256
+ assert flux_t5_sequence_length(_WordTokenizer(), [["a", _words(400)]]) == 512
+
+
+def test_no_tokenizer_keeps_the_pipeline_default():
+ assert flux_t5_sequence_length(None, ["a"]) is None
+
+
+def test_only_flux1_families_and_only_when_accepted():
+ pipe = types.SimpleNamespace(tokenizer_2 = _WordTokenizer())
+ params = {"max_sequence_length": None}
+ assert flux_t5_kwarg("flux.1", pipe, params, {"prompt": "a"}) == 256
+ assert flux_t5_kwarg("flux.1-kontext", pipe, params, {"prompt": "a"}) == 256
+ assert flux_t5_kwarg("flux.2-klein", pipe, params, {"prompt": "a"}) is None
+ assert flux_t5_kwarg("flux.1", pipe, {}, {"prompt": "a"}) is None
+ # An explicit value is never overridden.
+ assert flux_t5_kwarg("flux.1", pipe, params, {"prompt": "a", "max_sequence_length": 77}) is None
+
+
+def test_negative_counts_only_under_true_cfg():
+ pipe = types.SimpleNamespace(tokenizer_2 = _WordTokenizer())
+ params = {"max_sequence_length": None}
+ kw = {"prompt": "a", "negative_prompt": _words(300)}
+ assert flux_t5_kwarg("flux.1", pipe, params, kw) == 256
+ assert flux_t5_kwarg("flux.1", pipe, params, {**kw, "true_cfg_scale": 4.0}) == 512
+ # prompt_2 is what FLUX.1 feeds T5 when given.
+ assert flux_t5_kwarg("flux.1", pipe, params, {"prompt": "a", "prompt_2": _words(280)}) == 512
+
+
+def test_ideogram4_comfy_guidance_switches_where_sigma_falls_to_0_3():
+ from core.inference.diffusion_text_length import (
+ ideogram4_comfy_guidance_schedule,
+ ideogram4_sigmas,
+ )
+
+ sigmas = ideogram4_sigmas(20, 1024, 1024, 0.0, 1.75)
+ assert sigmas == sorted(sigmas, reverse = True) and sigmas[0] > 0.999 # clamped at logSNR -15
+ schedule = ideogram4_comfy_guidance_schedule(20, 1024, 1024)
+ assert schedule == [7.0] * 17 + [3.0] * 3
+ assert sigmas[16] > 0.3 >= sigmas[17]
diff --git a/studio/backend/tests/test_diffusion_more_families.py b/studio/backend/tests/test_diffusion_more_families.py
index 66a6700b490..286049f50d5 100644
--- a/studio/backend/tests/test_diffusion_more_families.py
+++ b/studio/backend/tests/test_diffusion_more_families.py
@@ -81,8 +81,8 @@ def test_flux1_krea_dev_is_trusted_non_gguf():
def test_flux1_krea_dev_generation_defaults():
# Model-card recipe: 28 steps at guidance 4.5. The generic "krea" key (Turbo's 8-step no-CFG shape) must NOT swallow it, and the krea-2 defaults must stay intact.
- assert default_generation_params("black-forest-labs/FLUX.1-Krea-dev") == (28, 4.5)
- assert default_generation_params("QuantStack/FLUX.1-Krea-dev-GGUF") == (28, 4.5)
+ assert default_generation_params("black-forest-labs/FLUX.1-Krea-dev") == (20, 3.5)
+ assert default_generation_params("QuantStack/FLUX.1-Krea-dev-GGUF") == (20, 3.5)
assert default_generation_params("krea/Krea-2-Turbo") == (8, 0.0)
assert default_generation_params("krea/Krea-2-Raw") == (52, 3.5)
@@ -108,7 +108,7 @@ def test_flux_dev_and_krea_do_not_inherit_the_schnell_nvfp4_checkpoint():
def test_flux2_klein_generation_defaults_distinguish_base_from_distilled():
for size in ("4B", "9B"):
- assert default_generation_params(f"unsloth/FLUX.2-klein-base-{size}") == (50, 4.0)
+ assert default_generation_params(f"unsloth/FLUX.2-klein-base-{size}") == (20, 5.0)
assert default_generation_params(f"unsloth/FLUX.2-klein-{size}") == (4, 1.0)
@@ -167,10 +167,10 @@ def test_prequant_exclusion_does_not_break_a_family_type_that_lacks_the_field():
def test_zimage_base_generation_defaults_are_not_the_distilled_recipe():
# The base is undistilled: 20 steps at guidance 4. The more specific "z-image-turbo" key sits
# ahead of "z-image", so the 9-step CFG-free Turbo recipe must not swallow it.
- assert default_generation_params("Tongyi-MAI/Z-Image") == (20, 4.0)
- assert default_generation_params("unsloth/Z-Image-GGUF") == (20, 4.0)
- assert default_generation_params("Tongyi-MAI/Z-Image-Turbo") == (9, 0.0)
- assert default_generation_params("unsloth/Z-Image-Turbo-GGUF") == (9, 0.0)
+ assert default_generation_params("Tongyi-MAI/Z-Image") == (25, 3.0)
+ assert default_generation_params("unsloth/Z-Image-GGUF") == (25, 3.0)
+ assert default_generation_params("Tongyi-MAI/Z-Image-Turbo") == (8, 0.0)
+ assert default_generation_params("unsloth/Z-Image-Turbo-GGUF") == (8, 0.0)
# ── lumina-2 family ──────────────────────────────────────────────────────────
@@ -346,8 +346,8 @@ def test_hidream_bf16_component_table_present():
def test_ideogram4_generation_defaults():
- # Model-card settings: 48 steps, guidance 7 (an exact match keeps the pipeline's recommended tapered schedule).
- assert default_generation_params("ideogram-ai/ideogram-4-fp8") == (48, 7.0)
+ # ComfyUI's template: 20 steps at constant guidance 7 (an explicit 48 / 7 still keeps the card's tapered schedule).
+ assert default_generation_params("ideogram-ai/ideogram-4-fp8") == (20, 7.0)
def test_ideogram4_bf16_reservation_table_present():
diff --git a/studio/backend/tests/test_diffusion_unified_edit.py b/studio/backend/tests/test_diffusion_unified_edit.py
index fcd53a6712b..d289bd40a34 100644
--- a/studio/backend/tests/test_diffusion_unified_edit.py
+++ b/studio/backend/tests/test_diffusion_unified_edit.py
@@ -139,9 +139,9 @@ def test_sampling_defaults_resolve_for_every_21_artifact_name():
"unsloth/Qwen-Image-2.1-GGUF",
"unsloth/Qwen-Image-2.1-FP8",
):
- assert default_generation_params(name) == (40, 1.0), name
+ assert default_generation_params(name) == (25, 1.0), name
# The generic key still owns the other Qwen-Image checkpoints.
- assert default_generation_params("Qwen/Qwen-Image-2512") == (20, 4.0)
+ assert default_generation_params("Qwen/Qwen-Image-2512") == (50, 4.0)
def test_explicit_edit_hands_ordered_rgba_images_and_explicit_geometry(backend21):
diff --git a/studio/backend/tests/test_openai_images_generations_route.py b/studio/backend/tests/test_openai_images_generations_route.py
index d054fcd2640..6ec61c01c76 100644
--- a/studio/backend/tests/test_openai_images_generations_route.py
+++ b/studio/backend/tests/test_openai_images_generations_route.py
@@ -36,12 +36,16 @@ from utils.api_errors import install_api_error_handlers
@pytest.mark.parametrize(
"repo_id, expected",
[
- ("unsloth/Z-Image-Turbo-GGUF", (9, 0.0)), # turbo entry, before the z-image fallback
- ("unsloth/Z-Image-GGUF", (20, 4.0)),
+ ("unsloth/Z-Image-Turbo-GGUF", (8, 0.0)), # turbo entry, before the z-image fallback
+ ("unsloth/Z-Image-GGUF", (25, 3.0)),
("unsloth/FLUX.1-schnell-GGUF", (4, 0.0)), # schnell entry, before the flux.1 entry
- ("black-forest-labs/FLUX.1-dev", (28, 3.5)),
+ ("black-forest-labs/FLUX.1-dev", (20, 3.5)),
("unsloth/FLUX.2-klein-4B-GGUF", (4, 1.0)),
- ("unsloth/Qwen-Image-2512-GGUF", (20, 4.0)),
+ ("unsloth/Qwen-Image-2512-GGUF", (50, 4.0)),
+ ("Qwen/Qwen-Image-Edit-2511", (40, 4.0)),
+ ("black-forest-labs/FLUX.1-Kontext-dev", (20, 2.5)),
+ ("black-forest-labs/FLUX.2-dev", (20, 4.0)),
+ ("stabilityai/stable-diffusion-xl-base-1.0", (25, 7.0)),
("some/unknown-model", (9, 0.0)), # fallback
("", (9, 0.0)),
],
@@ -60,14 +64,14 @@ def test_default_generation_params_specificity_ordering():
def test_default_generation_params_falls_back_to_base_repo():
# A local-path load: repo_id names no model, so the resolved base repo identifies it (and separates dev from schnell).
- assert default_generation_params("/models/my-ckpt", "black-forest-labs/FLUX.1-dev") == (28, 3.5)
+ assert default_generation_params("/models/my-ckpt", "black-forest-labs/FLUX.1-dev") == (20, 3.5)
assert default_generation_params("/models/my-ckpt", "black-forest-labs/FLUX.1-schnell") == (
4,
0.0,
)
assert default_generation_params("/models/my-ckpt", "Qwen/Qwen-Image") == (20, 4.0)
# repo_id wins when it already names the model; base repo is only a fallback.
- assert default_generation_params("unsloth/Z-Image-Turbo-GGUF", "Tongyi-MAI/Z-Image") == (9, 0.0)
+ assert default_generation_params("unsloth/Z-Image-Turbo-GGUF", "Tongyi-MAI/Z-Image") == (8, 0.0)
# Nothing identifiable -> fallback; None identifiers are skipped.
assert default_generation_params(None, None) == (9, 0.0)
assert default_generation_params("/models/x", None) == (9, 0.0)
@@ -232,12 +236,12 @@ def test_url_response_shape(client):
assert "url" in item and "b64_json" not in item # exclude_none drops the unused key
# Signed link, not the bearer-gated /file route: an OpenAI client downloads this URL with a plain GET and no auth header.
assert "/images/gallery/img0/file-signed?token=" in item["url"]
- # Z-Image-Turbo defaults (9 steps, 0 guidance) flow into the backend call.
+ # Z-Image-Turbo defaults (8 steps, 0 guidance) flow into the backend call.
assert client.backend.calls[0] == dict(
prompt = "a sloth",
width = 256,
height = 256,
- steps = 9,
+ steps = 8,
guidance = 0.0,
batch_size = 1,
expected_load = load_identity("unsloth/Z-Image-Turbo-GGUF", None, "z-image"),
@@ -298,7 +302,7 @@ def test_local_load_uses_base_repo_for_defaults(monkeypatch):
monkeypatch.setattr(gallery_module, "save", _save)
resp = cli.post("/v1/images/generations", json = {"prompt": "p", "size": "256x256"})
assert resp.status_code == 200
- assert backend.calls[0]["steps"] == 28 and backend.calls[0]["guidance"] == 3.5
+ assert backend.calls[0]["steps"] == 20 and backend.calls[0]["guidance"] == 3.5
def test_pipeline_runtime_error_is_sanitized_500(monkeypatch):
@@ -637,14 +641,14 @@ def test_generation_pins_the_status_read_it_derived_its_params_from(monkeypatch)
def test_replacement_retries_once_with_the_new_models_params(monkeypatch):
- # Z-Image-Turbo (9 steps, guidance 0) is replaced by Z-Image (20 steps, guidance 4). The first
+ # Z-Image-Turbo (8 steps, guidance 0) is replaced by Z-Image (25 steps, guidance 3). The first
# attempt is refused in-lock; the retry must re-derive from fresh state, not reuse the turbo's.
backend = _FakeBackend(replaced_by = [("unsloth/Z-Image-GGUF", None)])
cli, _ = _replacement_client(monkeypatch, backend)
resp = cli.post("/v1/images/generations", json = {"prompt": "p", "size": "256x256"})
assert resp.status_code == 200
assert len(backend.calls) == 1 # the refused attempt never generated
- assert (backend.calls[0]["steps"], backend.calls[0]["guidance"]) == (20, 4.0)
+ assert (backend.calls[0]["steps"], backend.calls[0]["guidance"]) == (25, 3.0)
def test_second_replacement_is_a_503_not_a_sanitized_500(monkeypatch):
diff --git a/studio/backend/tests/test_sd_cpp_backend.py b/studio/backend/tests/test_sd_cpp_backend.py
index 533c285f7fc..866c2fded01 100644
--- a/studio/backend/tests/test_sd_cpp_backend.py
+++ b/studio/backend/tests/test_sd_cpp_backend.py
@@ -754,6 +754,12 @@ def test_map_guidance_cfg_family_off_when_distilled():
assert _map_guidance(detect_family("qwen-image"), 4.0) == (4.0, None)
+def test_map_guidance_z_image_converts_diffusers_g_to_standard_cfg():
+ # The shared default is diffusers' g = 3 (ComfyUI cfg 4); sd.cpp's standard CFG must get 4, Turbo's 0 stays off.
+ assert _map_guidance(detect_family("Tongyi-MAI/Z-Image"), 3.0) == (4.0, None)
+ assert _map_guidance(detect_family("Tongyi-MAI/Z-Image-Turbo"), 0.0) == (1.0, None)
+
+
# ── status ────────────────────────────────────────────────────────────────────
diff --git a/studio/backend/tests/test_video_backend.py b/studio/backend/tests/test_video_backend.py
index fa21e258088..668989a84da 100644
--- a/studio/backend/tests/test_video_backend.py
+++ b/studio/backend/tests/test_video_backend.py
@@ -2203,12 +2203,12 @@ def test_wan_frame_snapping_4k_plus_1(fake_runtime):
def test_wan_ti2v_defaults_applied(fake_runtime):
- # No steps/guidance passed -> the Wan pipeline defaults (50 / 5.0).
+ # No steps/guidance passed -> ComfyUI's TI2V-5B template defaults (20 / 5.0).
backend = VideoBackend()
backend.load_pipeline("Wan-AI/Wan2.2-TI2V-5B-Diffusers", model_kind = "pipeline")
backend.generate(prompt = "a sloth")
call = backend._state.pipe.last_kwargs
- assert call["num_inference_steps"] == 50
+ assert call["num_inference_steps"] == 20
assert call["guidance_scale"] == 5.0
diff --git a/studio/backend/tests/test_video_families.py b/studio/backend/tests/test_video_families.py
index 26400463eaa..f03a1a3f444 100644
--- a/studio/backend/tests/test_video_families.py
+++ b/studio/backend/tests/test_video_families.py
@@ -240,13 +240,14 @@ def test_wan_snap_video_size_16():
def test_wan_generation_defaults():
- # Both Wan families default to the pipeline's 50 steps / CFG 5.0.
- assert default_video_generation_params(None, "Wan-AI/Wan2.2-TI2V-5B-Diffusers") == (50, 5.0)
- assert default_video_generation_params(None, "Wan-AI/Wan2.2-T2V-A14B-Diffusers") == (50, 5.0)
+ # ComfyUI's templates: TI2V-5B 20 steps / CFG 5, T2V-A14B 20 steps / CFG 3.5 (Lightning LoRA off).
+ assert default_video_generation_params(None, "Wan-AI/Wan2.2-TI2V-5B-Diffusers") == (20, 5.0)
+ assert default_video_generation_params(None, "Wan-AI/Wan2.2-T2V-A14B-Diffusers") == (20, 3.5)
+ assert default_video_generation_params("wan2.2-14b") == (20, 3.5)
# A GGUF filename carrying the family name still lands on the Wan defaults.
assert default_video_generation_params(
"wan2.2-ti2v-5b-Q4_K_M.gguf", "Wan-AI/Wan2.2-TI2V-5B-Diffusers"
- ) == (50, 5.0)
+ ) == (20, 5.0)
def test_generation_defaults_fallback_honors_family():
@@ -256,7 +257,7 @@ def test_generation_defaults_fallback_honors_family():
"/models/my-clip", "/models/my-clip", fallback = (50, 5.0)
) == (50, 5.0)
# A recognised token still wins over the fallback.
- assert default_video_generation_params("wan2.2-ti2v-5b", fallback = (8, 1.0)) == (50, 5.0)
+ assert default_video_generation_params("wan2.2-ti2v-5b", fallback = (8, 1.0)) == (20, 5.0)
def test_generation_defaults_wan_is_segment_not_substring():
@@ -268,8 +269,8 @@ def test_generation_defaults_wan_is_segment_not_substring():
"taiwan-clips.gguf", "user/taiwan-clips", fallback = (40, 4.0)
) == (40, 4.0)
# Genuine Wan identifiers (segment-initial, with a version suffix or separator) still match.
- assert default_video_generation_params("wan2.2-ti2v-5b-Q4_K_M.gguf") == (50, 5.0)
- assert default_video_generation_params(None, "Wan-AI/Wan2.2-T2V-A14B") == (50, 5.0)
+ assert default_video_generation_params("wan2.2-ti2v-5b-Q4_K_M.gguf") == (20, 5.0)
+ assert default_video_generation_params(None, "Wan-AI/Wan2.2-T2V-A14B") == (20, 3.5)
# An "ltxv" style name still resolves to LTX (trailing letters stay free).
assert default_video_generation_params("ltxv-2.3-distilled") == (8, 1.0)
assert default_video_generation_params("Lightricks/LTXV-2.3") == (40, 4.0)
@@ -332,10 +333,10 @@ def test_hv15_detection_and_flags():
def test_hv15_generation_defaults():
- # The community repacks ship a guider with guidance_scale 6.0 and the pipeline's own 50-step schedule.
+ # The community repacks ship a guider with guidance_scale 6.0; ComfyUI's template samples 20 steps.
assert default_video_generation_params(
None, "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v"
- ) == (50, 6.0)
+ ) == (20, 6.0)
def test_hv15_720p_checkpoints_never_route_to_the_480p_family():
diff --git a/studio/frontend/src/features/images/image-generation-defaults.ts b/studio/frontend/src/features/images/image-generation-defaults.ts
index 8f440c27fbb..8fd65115e34 100644
--- a/studio/frontend/src/features/images/image-generation-defaults.ts
+++ b/studio/frontend/src/features/images/image-generation-defaults.ts
@@ -11,35 +11,37 @@ const MODEL_DEFAULTS: Array<{
steps: number;
guidance: number;
}> = [
- { match: "z-image-turbo", steps: 9, guidance: 0 },
+ { match: "z-image-turbo", steps: 8, guidance: 0 },
// Krea 2 Raw is the undistilled base, so it must precede the distilled Krea key below.
{ match: "krea-2-raw", steps: 52, guidance: 3.5 },
{ match: "krea-2", steps: 8, guidance: 0 },
{ match: "flux.1-schnell", steps: 4, guidance: 0 },
// Kontext and the Krea dev finetune must precede generic FLUX.1.
- { match: "kontext", steps: 28, guidance: 2.5 },
- { match: "flux.1-krea", steps: 28, guidance: 4.5 },
- { match: "flux.1", steps: 28, guidance: 3.5 },
+ { match: "kontext", steps: 20, guidance: 2.5 },
+ { match: "flux.1-krea", steps: 20, guidance: 3.5 },
+ { match: "flux.1", steps: 20, guidance: 3.5 },
// Klein base is undistilled. The generic key below covers both distilled sizes.
- { match: "flux.2-klein-base", steps: 50, guidance: 4 },
+ { match: "flux.2-klein-base", steps: 20, guidance: 5 },
{ match: "flux.2-klein", steps: 4, guidance: 1 },
- { match: "flux.2-dev", steps: 28, guidance: 4 },
+ { match: "flux.2-dev", steps: 20, guidance: 4 },
// Qwen-Image-2.1 and its aliases, before the generic key.
- { match: "qwen-image-2.1", steps: 40, guidance: 1 },
- { match: "qwen-image-21", steps: 40, guidance: 1 },
- { match: "qwen_image_21", steps: 40, guidance: 1 },
- { match: "qwenimage21", steps: 40, guidance: 1 },
+ { match: "qwen-image-2.1", steps: 25, guidance: 1 },
+ { match: "qwen-image-21", steps: 25, guidance: 1 },
+ { match: "qwen_image_21", steps: 25, guidance: 1 },
+ { match: "qwenimage21", steps: 25, guidance: 1 },
+ { match: "qwen-image-edit", steps: 40, guidance: 4 },
+ { match: "qwen-image-2512", steps: 50, guidance: 4 },
{ match: "qwen-image", steps: 20, guidance: 4 },
- { match: "z-image", steps: 20, guidance: 4 },
- { match: "ideogram", steps: 48, guidance: 7 },
+ { match: "z-image", steps: 25, guidance: 3 },
+ { match: "ideogram", steps: 20, guidance: 7 },
{ match: "lumina", steps: 50, guidance: 4 },
{ match: "hunyuanimage", steps: 50, guidance: 3.25 },
{ match: "hidream-i1-dev", steps: 28, guidance: 0 },
{ match: "hidream-i1-fast", steps: 16, guidance: 0 },
{ match: "hidream", steps: 50, guidance: 5 },
{ match: "sdxl-turbo", steps: 3, guidance: 0 },
- { match: "stable-diffusion-xl", steps: 30, guidance: 7 },
- { match: "sdxl", steps: 30, guidance: 7 },
+ { match: "stable-diffusion-xl", steps: 25, guidance: 7 },
+ { match: "sdxl", steps: 25, guidance: 7 },
];
export function defaultsFor(repoId: string): {
diff --git a/studio/frontend/src/features/video/video-page.tsx b/studio/frontend/src/features/video/video-page.tsx
index ea22585783c..326fe912827 100644
--- a/studio/frontend/src/features/video/video-page.tsx
+++ b/studio/frontend/src/features/video/video-page.tsx
@@ -252,10 +252,13 @@ const MODEL_DEFAULTS: Array<{ match: string; steps: number; guidance: number }>
// "distilled" before the generic "ltx": the distilled model runs at 8 steps, guidance 1.
{ match: "distilled", steps: 8, guidance: 1 },
{ match: "ltx", steps: 40, guidance: 4 },
- // Wan2.2 pipelines default to 50 steps at CFG 5.0 (diffusers 0.39). The backend supplies the fps per family.
- { match: "wan", steps: 50, guidance: 5 },
- // HunyuanVideo-1.5 runs 50 steps; guidance 6 matches the guider the repo ships.
- { match: "hunyuanvideo", steps: 50, guidance: 6 },
+ // ComfyUI's templates for the same models. Wan2.2-T2V-A14B: 20 steps at CFG 3.5, before the generic Wan key.
+ { match: "a14b", steps: 20, guidance: 3.5 },
+ { match: "wan2.2-14b", steps: 20, guidance: 3.5 },
+ // Wan2.2-TI2V-5B: 20 steps at CFG 5. The backend supplies the fps per family.
+ { match: "wan", steps: 20, guidance: 5 },
+ // HunyuanVideo-1.5 runs 20 steps; guidance 6 matches the guider the repo ships.
+ { match: "hunyuanvideo", steps: 20, guidance: 6 },
];
function defaultsFor(repoId: string): { steps: number; guidance: number } {
diff --git a/studio/frontend/tests/image-generation-defaults.test.ts b/studio/frontend/tests/image-generation-defaults.test.ts
index 5693f061710..78d115c528b 100644
--- a/studio/frontend/tests/image-generation-defaults.test.ts
+++ b/studio/frontend/tests/image-generation-defaults.test.ts
@@ -17,8 +17,8 @@ import { readSrc } from "./helpers/kit.ts";
test("distinguishes Klein base checkpoints from distilled checkpoints", () => {
for (const size of ["4B", "9B"]) {
assert.deepEqual(defaultsFor(`unsloth/FLUX.2-klein-base-${size}`), {
- steps: 50,
- guidance: 4,
+ steps: 20,
+ guidance: 5,
});
assert.deepEqual(defaultsFor(`unsloth/FLUX.2-klein-${size}`), {
steps: 4,
@@ -33,7 +33,7 @@ test("keeps the existing family defaults and fallback", () => {
guidance: 3.5,
});
assert.deepEqual(defaultsFor("black-forest-labs/FLUX.1-dev"), {
- steps: 28,
+ steps: 20,
guidance: 3.5,
});
assert.deepEqual(defaultsFor("local/unknown-image-model"), {
@@ -158,3 +158,22 @@ test("every images-page canvas seed passes the quant provenance", () => {
assert.match(call.slice(0, 400), /transformerQuantSource: status\??\.resolved\?\.transformer_quant\?\.source/);
}
});
+
+test("defaults follow ComfyUI's official templates for the same model", () => {
+ for (const [id, want] of [
+ ["black-forest-labs/FLUX.1-dev", { steps: 20, guidance: 3.5 }],
+ ["black-forest-labs/FLUX.1-Krea-dev", { steps: 20, guidance: 3.5 }],
+ ["black-forest-labs/FLUX.1-Kontext-dev", { steps: 20, guidance: 2.5 }],
+ ["black-forest-labs/FLUX.2-dev", { steps: 20, guidance: 4 }],
+ ["Qwen/Qwen-Image-Edit-2511", { steps: 40, guidance: 4 }],
+ ["Qwen/Qwen-Image-2512", { steps: 50, guidance: 4 }],
+ ["Qwen/Qwen-Image", { steps: 20, guidance: 4 }],
+ ["Tongyi-MAI/Z-Image-Turbo", { steps: 8, guidance: 0 }],
+ // diffusers Z-Image guidance g equals ComfyUI cfg g + 1.
+ ["Tongyi-MAI/Z-Image", { steps: 25, guidance: 3 }],
+ ["stabilityai/stable-diffusion-xl-base-1.0", { steps: 25, guidance: 7 }],
+ ["ideogram-ai/ideogram-4-fp8", { steps: 20, guidance: 7 }],
+ ] as const) {
+ assert.deepEqual(defaultsFor(id), want, id);
+ }
+});
diff --git a/studio/frontend/tests/image-unified-edit.test.ts b/studio/frontend/tests/image-unified-edit.test.ts
index d1485b96f38..6bd27846eef 100644
--- a/studio/frontend/tests/image-unified-edit.test.ts
+++ b/studio/frontend/tests/image-unified-edit.test.ts
@@ -200,16 +200,16 @@ test("prompt helpers edit the visible instruction idempotently", () => {
);
});
-test("Qwen-Image-2.1 samples at 40 steps without guidance under every artifact name", () => {
+test("Qwen-Image-2.1 samples at 25 steps without guidance under every artifact name", () => {
for (const id of [
"Qwen/Qwen-Image-2.1",
"unsloth/Qwen-Image-2.1-GGUF",
"unsloth/Qwen-Image-2.1-FP8",
]) {
- assert.deepEqual(defaultsFor(id), { steps: 40, guidance: 1 }, id);
+ assert.deepEqual(defaultsFor(id), { steps: 25, guidance: 1 }, id);
}
assert.deepEqual(defaultsFor("Qwen/Qwen-Image-2512"), {
- steps: 20,
+ steps: 50,
guidance: 4,
});
});