Skip to content

Free, open-source AI video generator for Mac. Run LTX-2.5 and MiniMax H3 (Hailuo) locally on Apple Silicon: text-to-video and image-to-video with sound, lip-sync, character LoRA training, songs from lyrics (YuE2) and image editing (Qwen-Image-Edit). One-click install via Pinokio. Offline: no cloud, no API key, no NVIDIA GPU.

feat(image-studio): standardize Qwen + Flux tiers — Fast / Medium / Quality

mrbizarro35205a3c8ec8
diff --git a/mlx_ltx_panel.py b/mlx_ltx_panel.py
index 0f08e03..77ee22f 100644
--- a/mlx_ltx_panel.py
+++ b/mlx_ltx_panel.py
@@ -5331,23 +5331,20 @@ def _build_image_engine_config(engine_override: str) -> agent_image_engine.Image
engine_override = (engine_override or "auto").lower()
if engine_override == "auto":
return _load_agent_image_config()
- if engine_override == "qwen_edit_inline":
- # Iteration default: Qwen-Image-Edit-2511 (Apache 2.0, ungated)
- # at Q6 + 8 steps. 2511 is the current best Qwen-Edit — stronger
- # character consistency than 2509, mitigated multi-ref drift,
- # built-in popular LoRAs. Q6 (vs prior Q4) is the Apple-Silicon
- # community sweet spot — ~4-6% quality loss vs full precision
- # instead of 8-12% for Q4, with negligible speed impact on
- # M4 Max 64 GB.
- return agent_image_engine.ImageEngineConfig(
- kind="mflux", mflux_model="Qwen/Qwen-Image-Edit-2511",
- mflux_family="qwen_edit", mflux_quantize=6, mflux_steps=8,
- )
+ # ===== Qwen-Image-Edit-2511 — Fast / Medium / Quality ====================
+ # All three share the same ~24 GB Qwen-Image-Edit-2511 model on disk
+ # (Apache 2.0, ungated). The split is steps × quantization × LoRA.
+ # 2511 is the current best Qwen-Edit — stronger character consistency
+ # than 2509 and mitigated multi-ref drift.
+ # Q6 (vs prior Q4) is the Apple-Silicon community sweet spot — ~4-6%
+ # quality loss vs full precision instead of 8-12% for Q4, with
+ # negligible speed impact on M4 Max 64 GB.
if engine_override == "qwen_edit_lightning_inline":
- # Fast preset: 2511 + Q6 + 4 steps + Lightning distillation LoRA.
- # Pin the bf16 4-step file via the `repo:filename` collection-format
- # syntax so mflux doesn't pull all four LoRA variants and stall.
- # Prior config had a 2509 base + 2511 Lightning version mismatch.
+ # Qwen FAST — 4-step Lightning distillation LoRA, Q6. The
+ # `repo:filename` collection-format syntax pins the bf16 4-step
+ # variant so mflux doesn't pull all four LoRA versions and stall.
+ # Lightning was distilled at scale 1.0 — don't change it.
+ # Total wall-time ~1:30 per image (50 s cold + 40 s denoise).
return agent_image_engine.ImageEngineConfig(
kind="mflux", mflux_model="Qwen/Qwen-Image-Edit-2511",
mflux_family="qwen_edit", mflux_quantize=6, mflux_steps=4,
@@ -5356,43 +5353,63 @@ def _build_image_engine_config(engine_override: str) -> agent_image_engine.Image
],
mflux_lora_scales=[1.0],
)
+ if engine_override == "qwen_edit_inline":
+ # Qwen MEDIUM — 8-step Q6, no LoRA, no CFG. Sits between Fast's
+ # distilled output (sometimes over-smooth) and Quality's 40-step
+ # cost. ~2:35 per image (50 s cold + 105 s denoise).
+ return agent_image_engine.ImageEngineConfig(
+ kind="mflux", mflux_model="Qwen/Qwen-Image-Edit-2511",
+ mflux_family="qwen_edit", mflux_quantize=6, mflux_steps=8,
+ )
if engine_override == "qwen_edit_high_inline":
- # Final-render preset: 2511 Q8 + 40 steps. 40 steps matches the
- # official Diffusers reference (qwen.ai blog: true_cfg_scale=4.0
- # + guidance_scale=1.0 + steps=40). mflux's --guidance maps to
- # true_cfg_scale on the qwen-edit CLI.
+ # Qwen QUALITY — 40-step Q8 + CFG 4.0. Matches the Qwen
+ # reference recipe (qwen.ai blog: true_cfg_scale=4.0 +
+ # guidance_scale=1.0 + steps=40). mflux's --guidance maps to
+ # true_cfg_scale on the qwen-edit CLI. ~5 min per image
+ # (60 s cold + 240 s denoise).
return agent_image_engine.ImageEngineConfig(
kind="mflux", mflux_model="Qwen/Qwen-Image-Edit-2511",
mflux_family="qwen_edit", mflux_quantize=8, mflux_steps=40,
mflux_guidance=4.0,
)
- if engine_override == "flux2_inline":
- # Q6 instead of Q4 — same Apple-Silicon sweet-spot rationale as
- # the qwen presets above.
- return agent_image_engine.ImageEngineConfig(
- kind="mflux", mflux_model="Runpod/FLUX.2-klein-4B-mflux-4bit",
- mflux_family="flux2", mflux_quantize=6,
- )
+ # ===== FLUX.2 family + Z-Image-Turbo ====================================
+ # Flux ships TWO edit tiers (no middle): the distilled Klein-Edit
+ # (fast, illustrative aesthetic) and the non-distilled Klein-Base-Edit
+ # (slower, photographic). A Medium tier would require a separate
+ # distilled-with-CFG checkpoint that doesn't exist upstream; we keep
+ # the dropdown honest at Fast / Quality only.
if engine_override == "flux2_edit_inline":
- # FLUX.2 Klein-Edit (distilled klein-4b). mflux's flux2_edit_generate.py
- # FORCES guidance == 1.0 on distilled bases. Q6 default for the
- # same reason as above (Q4 → 8-12% quality loss; Q6 → 4-6%).
+ # Flux FAST — Klein-Edit (distilled klein-4b), 4-step Q6. mflux's
+ # flux2_edit_generate.py FORCES guidance=1.0 on distilled bases.
+ # Output has the distinctive illustrative / cartoonish look that's
+ # a feature for stylized work and a downside for photoreal. Q6
+ # over Q4 buys ~4-6 percentage points of quality. ~50 s per image
+ # (30 s cold + 20 s denoise).
return agent_image_engine.ImageEngineConfig(
kind="mflux", mflux_model="flux2-klein-4b",
mflux_family="flux2_edit", mflux_quantize=6,
mflux_steps=4, mflux_guidance=1.0,
)
if engine_override == "flux2_edit_high_inline":
- # FLUX.2 Klein-Base-Edit (NON-distilled klein-base-4b). Real CFG,
- # ~3-5 min/image at Q8 + 25 steps. Photographic output (vs the
- # illustrative look of distilled flux2_edit). Already at Q8 — no
- # change.
+ # Flux QUALITY — Klein-Base-Edit (NON-distilled klein-base-4b),
+ # 25-step Q8 + CFG 4.0. Real CFG = photographic output (vs the
+ # illustrative look of distilled flux2_edit). Already at Q8.
+ # ~4 min per image (30 s cold + 210 s denoise).
return agent_image_engine.ImageEngineConfig(
kind="mflux", mflux_model="flux2-klein-base-4b",
mflux_family="flux2_edit", mflux_quantize=8,
mflux_steps=25, mflux_guidance=4.0,
)
+ if engine_override == "flux2_inline":
+ # Flux2 T2I — Klein 4B, 4-step Q6. No-refs path. Same Q6 sweet-
+ # spot rationale as qwen. ~24 s per image (15 s cold + 9 s denoise).
+ return agent_image_engine.ImageEngineConfig(
+ kind="mflux", mflux_model="Runpod/FLUX.2-klein-4B-mflux-4bit",
+ mflux_family="flux2", mflux_quantize=6,
+ )
if engine_override == "z_image_turbo_inline":
+ # Z-Image-Turbo — compact T2I, 9-step Q4 distilled. Fastest pure
+ # T2I in the panel. ~17 s per image (12 s cold + 5 s denoise).
return agent_image_engine.ImageEngineConfig(
kind="mflux", mflux_model="filipstrand/Z-Image-Turbo-mflux-4bit",
mflux_family="z_image_turbo", mflux_quantize=4,
@@ -6420,21 +6437,30 @@ class Handler(BaseHTTPRequestHandler):
# PER candidate. cold_start_sec is the one-time subprocess
# model-load that's paid ONCE per batch, regardless of n.
# The JS estimator does: total = cold_start + n × sec_per_image.
+ #
+ # All numbers anchored to measured wall-time on M4 Max @ 1024
+ # for mflux engines, 2560×1440 for HiDream. mflux subprocess
+ # cold loads are not negligible: Qwen-Edit Q6 = ~50s,
+ # Qwen-Edit Q8 = ~60s, Flux2 family ~30s, Z-Image-Turbo ~12s.
+ # The previous table folded loads into sec_per_image, which
+ # made N=1 understate and N=4+ overstate the wall-time. The
+ # studio now estimates correctly across all batch sizes.
ENGINES = [
- # mflux-family engines load the model once per subprocess
- # too, but their loads are quick (~15s) so we fold them
- # into sec_per_image rather than expose the split.
- ("qwen_edit_lightning_inline", "Qwen/Qwen-Image-Edit-2511", 24.0, 13.0, 0.0),
- ("qwen_edit_inline", "Qwen/Qwen-Image-Edit-2511", 24.0, 60.0, 0.0),
- ("qwen_edit_high_inline", "Qwen/Qwen-Image-Edit-2511", 24.0, 300.0, 0.0),
- ("flux2_edit_inline", None, 0.0, 30.0, 0.0),
- ("flux2_edit_high_inline", None, 0.0, 240.0, 0.0),
- ("flux2_inline", "Runpod/FLUX.2-klein-4B-mflux-4bit", 4.0, 12.0, 0.0),
- ("z_image_turbo_inline", "filipstrand/Z-Image-Turbo-mflux-4bit", 3.0, 6.0, 0.0),
+ # Qwen-Image-Edit-2511 — three-tier ladder. ~24 GB
+ # download (one-time) shared across all three tiers.
+ ("qwen_edit_lightning_inline", "Qwen/Qwen-Image-Edit-2511", 24.0, 40.0, 50.0),
+ ("qwen_edit_inline", "Qwen/Qwen-Image-Edit-2511", 24.0, 105.0, 50.0),
+ ("qwen_edit_high_inline", "Qwen/Qwen-Image-Edit-2511", 24.0, 240.0, 60.0),
+ # FLUX.2 Klein edit family — distilled vs base-edit.
+ ("flux2_edit_inline", None, 0.0, 20.0, 30.0),
+ ("flux2_edit_high_inline", None, 0.0, 210.0, 30.0),
+ # Flux2 T2I + Z-Image-Turbo (no refs). Both have small
+ # cold loads (bundled mflux engines).
+ ("flux2_inline", "Runpod/FLUX.2-klein-4B-mflux-4bit", 4.0, 9.0, 15.0),
+ ("z_image_turbo_inline", "filipstrand/Z-Image-Turbo-mflux-4bit", 3.0, 5.0, 12.0),
# HiDream lab subprocess: ~45s cold load (BF16 weights →
- # MLX) is real and material to the estimate for n=1
- # requests. Denoise times are the measured per-candidate
- # wall-time from the May 2026 step+FBCache bench at HD.
+ # MLX) per batch. Denoise times measured from the May 2026
+ # step+FBCache bench at HD 2560×1440.
("hidream_fast_inline", None, 0.0, 45.0, 45.0),
("hidream_inline", None, 0.0, 80.0, 45.0),
("hidream_quality_inline", None, 0.0, 120.0, 45.0),
@@ -18493,16 +18519,16 @@ HTML = r"""<!doctype html>
<div class="studio-engine-row">
<select id="imgStudioEngine" onchange="imgStudioUpdateValidity();imgStudioRefreshEngineStatus();imgStudioUpdateEstimate();if(typeof renderLorasList==='function')renderLorasList()">
<option value="auto">Auto (use Settings)</option>
- <option value="qwen_edit_lightning_inline" selected>Qwen-Image-Edit-2511 + Lightning &mdash; FAST (multi-ref &middot; 4-step Q6, ~90 s/image @ 1024)</option>
- <option value="qwen_edit_inline">Qwen-Image-Edit-2511 &mdash; polished (multi-ref &middot; 8-step Q6, ~2:35 /image @ 1024)</option>
- <option value="qwen_edit_high_inline">Qwen-Image-Edit-2511 &mdash; best (multi-ref &middot; 40-step Q8, ~5 min/image @ 1024)</option>
- <option value="flux2_edit_inline">FLUX.2 Klein-Edit fast (multi-ref &middot; 4-step Q6, ~30 s/image)</option>
- <option value="flux2_edit_high_inline">FLUX.2 Klein-Base-Edit photoreal (multi-ref &middot; 25-step Q8, ~3-5 min/image)</option>
- <option value="flux2_inline">FLUX.2 [klein] 4B (fast T2I, no refs)</option>
- <option value="z_image_turbo_inline">Z-Image-Turbo (compact T2I, no refs)</option>
- <option value="hidream_fast_inline">HiDream Fast (3-step &middot; ~45 s preview, slight bg softness, 18 GB)</option>
- <option value="hidream_inline">HiDream Medium (6-step + FBCache &middot; ~80 s, character-preserving, 18 GB)</option>
- <option value="hidream_quality_inline">HiDream Quality (12-step + light FBCache &middot; ~120 s, best detail, 18 GB)</option>
+ <option value="qwen_edit_lightning_inline" selected>Qwen Fast (Lightning &middot; 4-step Q6, ~1:30 / image, multi-ref)</option>
+ <option value="qwen_edit_inline">Qwen Medium (8-step Q6, ~2:35 / image, multi-ref)</option>
+ <option value="qwen_edit_high_inline">Qwen Quality (40-step Q8 + CFG 4.0, ~5 min / image, multi-ref)</option>
+ <option value="flux2_edit_inline">Flux Fast (Klein-Edit 4-step Q6, ~50 s / image, multi-ref, illustrative)</option>
+ <option value="flux2_edit_high_inline">Flux Quality (Klein-Base-Edit 25-step Q8 + CFG 4.0, ~4 min / image, multi-ref, photoreal)</option>
+ <option value="flux2_inline">Flux2 T2I (Klein 4B 4-step Q6, ~24 s / image, no refs)</option>
+ <option value="z_image_turbo_inline">Z-Image Turbo (T2I, ~17 s / image, compact, no refs)</option>
+ <option value="hidream_fast_inline">HiDream Fast (3-step, ~3:45 / 4-img batch, slight bg softness)</option>
+ <option value="hidream_inline">HiDream Medium (6-step + FBCache, ~6 min / 4-img batch, character-preserving)</option>
+ <option value="hidream_quality_inline">HiDream Quality (12-step + light FBCache, ~9 min / 4-img batch, best detail)</option>
<option value="mock_inline" id="imgStudioMockOption" hidden>Mock (testing — debug only)</option>
</select>
<span class="engine-status-pill" id="imgStudioEnginePill"