ComfyUI Node
H3 • Render Image
Sample one H3 latent frame and decode one image using Fizgig's temporal-group technique. Prompt/reference preparation stays cached when only seed or steps changes. No video packet is sampled.
H3 • Render Image
- image_job
- image
- run_info
◄steps50►
◄seed42►
◄sampler_nameer_sde►
◄schedulersimple►
CategoryMiniMax H3/Image Studio
Inputs (5)
| Name | Type | Default | Description |
|---|---|---|---|
| image_job | H3_STILL_JOB | Connect Text to Image or Image Edit. | |
| steps | INT | 501–200 | Always manual. Start at 50 without a LoRA, or 20 with the optional Larry v4 adapter at 0.38 (Fizgig's still recipe). Fewer steps trade quality for speed. |
| seed | INT | 420–18446744073709550000 | Seed for repeatable sampling. Changing only seed or steps reuses cached prompt/reference encoding. |
| sampler_nameopt | COMBO | er_sde | ER-SDE is the Fizgig still starting recipe. Change only for controlled comparisons. |
| scheduleropt | COMBO | simple | Simple is the starting schedule. Uses the loaded H3 model's native flow shifts. |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| image | IMAGE | Exactly one image. Connect directly to Save Image or Preview Image. |
| run_info | STRING | Actual canvas, steps, seed, sampler, reference count and sampling/decode seconds; excludes model loading and cached preparation. |