| model | MODEL | | MiniMax H3 diffusion model (e.g. from UNETLoader or a ⭐ Star Minimax All In One MODEL output). |
| clip | CLIP | | The MiniMax H3 text encoder (Qwen3-VL-32B, clip type 'minimax'). |
| vae | VAE | | The MiniMax H3 video VAE (e.g. minimax_h3_video_vae_fp16). |
| image | IMAGE | | — |
| mask | MASK | | 1/white marks the regions to regenerate. |
| prompt | STRING | | What to generate in the masked area. The source image is <Picture 1>, extra reference images are <Picture 2>, <Picture 3>, ... |
| reference | COMBO | masked source | How the source crop enters the conditioning as <Picture 1>: 'masked source' shows it with the inpaint area blanked to gray (the mask's hole signals what to fill), 'source' shows it untouched (edit style - the model can see what was there), 'off' sends no picture to the text encoder at all. |
| megapixels | COMBO | native | Render resolution of the inpaint crop. 'native' keeps the crop's own pixels; the MP presets rescale crop + mask to that pixel area first - H3 stills hit their quality sweet spot around 3 MP and up. The result is scaled back to the crop size for stitching. |
| seed | INT | 00–18446744073709550000 | — |
| steps | INT | 201–10000 | — |
| cfg | FLOAT | 1.00–100 | CFG scale. At 1.0 the negative prompt is skipped and the positive-only basic guider is used (the H3 preset). |
| sampler_name | COMBO | res_multistep | 44 options: euler, euler_cfg_pp, euler_ancestral, euler_ancestral_cfg_pp, heun, heunpp2, +38 |
| scheduler | COMBO | simple | 'beta' or 'normal' tends to outperform 'simple' for reference-heavy prompts. |
| denoise | FLOAT | 1.000–1 | — |
| negative_prompt | STRING | | Negative prompt - only encoded when cfg differs from 1.0. |
| expand_pixels | INT | 320–512 | Context margin added around the masked area before sampling - more context helps the model blend the fill into the scene. |
| mask_blur | FLOAT | 16.00–256 | Grow + blur applied to the denoise mask (same as the Flux2 inpainter's mask blur). Soft edges translate directly into per-row denoise strength on H3. |
| blend_pixels | FLOAT | 16.00–256 | Feather width of the blend mask used when stitching the decoded still back into the original image. |
| ref_image_size | COMBO | match | Reference image sizing. 'match' scales each ref (down only, keeping aspect) to the render area; 'max' uses a 2048px short edge for best identity fidelity but is slower. |
| ref_imagesopt | COMFY_AUTOGROW_V3 | | — |
| sigmasopt | SIGMAS | | Optional exact sigma schedule (e.g. from a '⭐ Star Custom Sigmas' node) - overrides the scheduler, steps and denoise widgets. |
| previewopt | STAR_PREVIEW | | Optional live sampling preview from a '⭐ Star Preview' node. |
