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): switch HiDream default to Q6 (2x faster, 30% less RAM)

mrbizarro8a489536de28
diff --git a/agent/image_engine.py b/agent/image_engine.py
index c364a39..d838a06 100644
--- a/agent/image_engine.py
+++ b/agent/image_engine.py
@@ -32,10 +32,11 @@ This module is the dispatch layer. Four backends ship today:
- **hidream** — fully-local HiDream-O1-Image-Dev (8B Qwen3-VL-based unified
pixel-patch transformer, MIT). Lives in its own venv outside Phosphene
at `/Users/salo/HIDREAM-O1-MLX-LAB-active/`. Subprocess pattern, mirrors
- mflux. Comfortable+ tier (32 GB+) due to ~11.5 GB working set at Q8,
- 1024×1024 in ~67s. Strong prompt fidelity, recognisable subjects across
- photo/anime/painting/macro/architecture; faint 32-pixel patch grid in
- flat regions (architectural, not fixable without retraining).
+ mflux. Default quant **Q6** (~8.5 GB working set, 1024×1024 in ~36s) —
+ bandwidth-bound on this hardware so Q6 is twice as fast as Q8 with
+ equivalent prompt fidelity. Q4 ships dark — never use. Strong prompt
+ fidelity across photo/anime/painting/macro/architecture; faint 32-pixel
+ patch grid in flat regions (architectural).
Recommended defaults (May 2026):
- **Comfortable+ (32 GB+)** → `Runpod/FLUX.2-klein-4B-mflux-4bit`
@@ -997,7 +998,11 @@ def _generate_bfl(prompt: str, n: int, width: int, height: int,
HIDREAM_LAB_DIR = Path("/Users/salo/HIDREAM-O1-MLX-LAB-active")
HIDREAM_DEFAULT_PY = HIDREAM_LAB_DIR / ".venv" / "bin" / "python"
-HIDREAM_DEFAULT_MODEL = HIDREAM_LAB_DIR / "mlx_models" / "hidream-o1-dev-q8"
+# Q6 sweet spot: 2x faster than Q8 (36s vs 67s per 1024×1024) and 30% less
+# RAM (8.5 GB vs 11.5 GB) with the same prompt fidelity. mlx-vlm's quantized
+# matmul kernels are bandwidth-bound and Q6 has fewer bits per weight.
+# Q4 ships dark (brightness collapses) — never use.
+HIDREAM_DEFAULT_MODEL = HIDREAM_LAB_DIR / "mlx_models" / "hidream-o1-dev-q6"
HIDREAM_GENERATE_SCRIPT = HIDREAM_LAB_DIR / "scripts" / "hidream_o1" / "generate_hidream_o1_mlx.py"
HIDREAM_PATCH_SIZE = 32 # HiDream operates on patch-aligned multiples; we round if asked for non-aligned
diff --git a/docs/IMAGE_GEN_RESEARCH_2026-05.md b/docs/IMAGE_GEN_RESEARCH_2026-05.md
index 26e77b0..35cda04 100644
--- a/docs/IMAGE_GEN_RESEARCH_2026-05.md
+++ b/docs/IMAGE_GEN_RESEARCH_2026-05.md
@@ -8,7 +8,7 @@ LTX 2.3's t2v is weak; image-gen quality is the dominant lever for video output.
- **"Flux Klein"** = **FLUX.2 [klein] 4B** by Black Forest Labs (Jan 2026). Apache 2.0, mflux-compatible at 4-bit (~4.3 GB), 4-step inference, ~12-18 s/image on Comfortable Mac. **Ship as default.**
- **mflux 0.17.x is now multi-family.** Legacy `mflux-generate` is FLUX.1-only. New CLIs: `mflux-generate-flux2`, `mflux-generate-z-image-turbo`, `mflux-generate-fibo`, `mflux-generate-qwen`, `mflux-generate-kontext`. Panel currently calls legacy → silently misses every new family.
- **Z-Image-Turbo (Tongyi/Alibaba, Nov 2025)** is the Compact-tier default — 6B model, ~5.9 GB at 4-bit, 8-9 steps, fits 16 GB.
-- **HiDream-O1-Image-Dev (May 2026 addition)** — 8B Qwen3-VL pixel-patch transformer, MIT licence, runs out of a separate lab venv (NOT mflux). Q8 ~10 GB, ~67 s / 1024×1024 on Comfortable+. Strong prompt fidelity across all genres tested. Available as `kind="hidream"` in `agent/image_engine.py` since commit 45cad69.
+- **HiDream-O1-Image-Dev (May 2026 addition)** — 8B Qwen3-VL pixel-patch transformer, MIT licence, runs out of a separate lab venv (NOT mflux). **Q6 default** (~8 GB, ~36 s / 1024×1024) — 2× faster than Q8 with equivalent fidelity (mlx is bandwidth-bound). Strong prompt fidelity across all genres tested. Available as `kind="hidream"` in `agent/image_engine.py` since commit 45cad69.
## The single default to ship
@@ -63,7 +63,7 @@ def _default_image_engine_for_tier(tier: str) -> dict:
| FIBO | `briaai/FIBO` | 8B | ~10-12 | Open-RAIL-ish | 30 |
| Qwen-Image | `filipstrand/Qwen-Image-mflux-6bit` | 20B | ~16 (6-bit) | Apache 2.0 | 25-50 |
| Flex.1-alpha | `ostris/Flex.1-alpha` | 8B | ~9-10 | Apache 2.0 | 25 |
-| **HiDream-O1-Image-Dev** | `HiDream-ai/HiDream-O1-Image-Dev` | 8B | **~10 (Q8)** | **MIT** | 28 |
+| **HiDream-O1-Image-Dev** | `HiDream-ai/HiDream-O1-Image-Dev` | 8B | **~8 (Q6)** | **MIT** | 28 |
## mflux 0.17.x per-family CLIs
@@ -105,14 +105,14 @@ The lab port lives outside Phosphene at `/Users/salo/HIDREAM-O1-MLX-LAB-active/`
- **Patch-grid artifact** in flat regions (sky, water, walls). Architectural — not fixable without retraining or an overlap-blend post-pass.
- **No edit/multi-ref yet.** Architecture supports it; lab pipeline is text-to-image only as of May 2026. Refs continue to flow through `mflux qwen-edit`.
-### Tier guidance
+### Tier guidance (revised May 2026 after Q6 measurement)
| Tier | RAM | Use HiDream? |
|---|---|---|
-| Compact (16-32 GB) | tight | NO — not enough RAM headroom alongside LTX. Use Z-Image-Turbo. |
-| Comfortable (32-79 GB) | OK | OPTIONAL — slower than klein but better fidelity for hero shots. |
-| Roomy (80-119 GB) | YES | Routine. |
-| Studio (120 GB+) | YES | Routine; consider 2048×2048 for finals. |
+| Compact (16-32 GB) | tight | OPTIONAL at Q6 (8.5 GB) — fits but no LTX headroom. Z-Image-Turbo still safer. |
+| Comfortable (32-79 GB) | OK | YES at Q6 — 36s vs Z-Image-Turbo's 80s, lower deterministic RAM. |
+| Roomy (80-119 GB) | YES | Routine; pick Q6 for speed or Q8 for max safety margin. |
+| Studio (120 GB+) | YES | Routine; consider 2048×2048 for finals (~5 min at Q8). |
## Implementation plan (20 ranked items)
diff --git a/mlx_ltx_panel.py b/mlx_ltx_panel.py
index 8636e67..221abb4 100644
--- a/mlx_ltx_panel.py
+++ b/mlx_ltx_panel.py
@@ -5552,7 +5552,7 @@ class Handler(BaseHTTPRequestHandler):
("flux2_edit_high_inline", None, 0.0, 240.0),
("flux2_inline", "Runpod/FLUX.2-klein-4B-mflux-4bit", 4.0, 12.0),
("z_image_turbo_inline", "filipstrand/Z-Image-Turbo-mflux-4bit", 3.0, 6.0),
- ("hidream_inline", None, 0.0, 67.0),
+ ("hidream_inline", None, 0.0, 36.0),
("mock_inline", None, 0.0, 0.5),
]
out = []
@@ -15086,7 +15086,7 @@ HTML = r"""<!doctype html>
<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_inline">HiDream-O1-Image-Dev Q8 (high-fidelity T2I, no refs &middot; ~67 s/image, 11 GB) [NEW]</option>
+ <option value="hidream_inline">HiDream-O1-Image-Dev Q6 (high-fidelity T2I, no refs &middot; ~36 s/image, 8.5 GB) [NEW]</option>
<option value="mock_inline" id="imgStudioMockOption" hidden>Mock (testing — debug only)</option>
</select>
<span class="engine-status-pill" id="imgStudioEnginePill"