Adds HiDream-O1-Image-Dev-2604 fp8 as the 7th Studio lane and the
quality leader for general/photoreal stills (Artificial Analysis #1
single-model open-weight T2I, Elo 1189). A 9B pixel-level unified
transformer (Qwen3-VL backbone, no separate VAE/text-encoder); takes a
rich natural-language prompt; Dev-2604 is distilled (28-step, CFG-off).
- workflows/hidream_o1.json: Loader -> Conditioning -> Sampler -> Save,
model_type=auto (dev_2604 fp8), native 2048^2.
- build_studio_pipe.py (v0.12.0): hidream lane, DIRECTOR_HIDREAM_SYS
(prose art-director), _comfy_hidream, routing/label/valves. _await_output
now skips type=temp images so the gallery-served SaveImage png is
returned (HiDream's sampler emits a temp preview).
- HiDream-O1 has NO native ComfyUI support: entrypoint clones the
third-party Saganaki22/HiDream_O1-ComfyUI node (source-reviewed: no
subprocess/eval/network beyond an opt-in HF downloader) + an idempotent
transformers-5.x compat patch (create_causal_mask input_embeds->
inputs_embeds, drop cache_position) so the model runs on our image's
transformers 5.11.
- download_hidream_o1.sh fetches the ~8.8 GB weights folder.
Renders at native 2048^2 (~15 GB GPU0), ~3-4 min cold / ~44 s warm on a
3090 (sdpa attn; flash-attn not built for sm_86). Validated live: clean
2048^2 png to the gallery.
Co-authored-by: noonghunna <[email protected]>
Co-authored-by: Claude Opus 4.8 <[email protected]>
