MiniMax H3 Image • Image to Image

Image-to-image with H3 is where the "video model generating stills" thing gets genuinely weird, and H3ImageToImagePrepare - "MiniMax H3 Image • Image to Image" - exists to hide the weirdness from you. The short version: for multi-frame editing, H3 anchors your source image as frame 0 of the temporal packet (that's the FL2VA keyframe path), and the model edits around it. For a true one-frame edit, that anchor is impossible - the frame-0 keyframe would be the only output frame - so the node silently switches to REF2VA-style source conditioning instead. You just pick a profile and it handles the rest.

Inputs worth understanding

The inputs are the normal prepare-node lineup: clip (the Qwen text/vision encoder), vae (the H3 video VAE for multi-frame, or the experimental image VAE for one-frame), source_image, an edit_instruction, and width/height for the generation canvas (the source is fitted to it before encoding). The ones worth understanding:

• quality_profile - recommended | 5 frames is the default. One-frame (single image | 1 frame (image VAE)) switches to the REF2VA source-conditioning path, so it expects a hybrid or REF2VA checkpoint plus the H3 image VAE. Multi-frame profiles keep the FL2VA anchor.
• source_fidelity (default 0.75) - controls how strongly the still prompt asks H3 to preserve identity, pose, composition, and geometry. It is not denoise strength and doesn't alter the sigma schedule. This is the single most-misunderstood knob in the pack.
• source_fit - crop_center (default), contain_pad, or stretch for fitting the source to the canvas before VAE encoding.
• optimize_for_still (default on) - wraps your instruction in a locked-camera still-image prompt and adds source-preservation language. Sampling settings are untouched.

Outputs

Outputs: positive conditioning for the sampler's positive input, h3_latent (which carries the anchored frame-0 keyframe in multi-frame mode), fitted_source for preview/comparison, requested_frames to feed Exact Frame Decode, image_prompt (the final optimized edit prompt), and run_info with the temporal packet and checkpoint expectations.

Practical stuff

The practical stuff first: in every image-to-image workflow you must select a file in each Load Image node, or you'll get a "Load Image - image is missing" error - that's the most common support question for this pack, and it's not a bug, it's an empty input. Second, match your source image to the generation canvas for lower preprocessing cost; crop_center on a wildly different aspect ratio is where you lose composition you didn't mean to lose.

One design note that shows the pack is carefully thought out: the multi-frame output strategy is set to first_stable_edit, so the decoder's recommended_index doesn't hand you the unchanged source anchor. That was a real bug in earlier versions (short image-to-image packets selecting the untouched anchor as the default output) and it's fixed in the current decoder - which is also why the README insists on a full restart after updating, not just a refresh, if you see stale behavior.

Install

Install is the pack: ComfyUI Manager (search MiniMax H3 Image Studio) or git clone https://github.com/astropuzzo/ComfyUI-MiniMax-H3-Image-Studio.git into custom_nodes, then restart. ComfyUI 0.30.0+ and the H3 model stack from the official ComfyUI MiniMax H3 guide; the audio VAE is not required for image output.