MiniMax H3 Image • Advanced Combined Prepare

If you've seen the Text to Image, Image to Image, and Reference Edit nodes in this pack and thought "that's three nodes I'll have to learn," H3ImagePrepare - "MiniMax H3 Image • Advanced Combined Prepare" - is the consolidation. It does all three conditioning paths in a single node, and it adds up to nine reference inputs on top. It's the one to reach for when you want a Swiss-army workflow or you're not sure yet which mode a job needs.

The mode switcher

The mode dropdown picks the path. Text to image and image to image use FL2VA; reference edit uses REF2VA. From there the node does the same heavy lifting every prepare node in this pack does: it rounds your canvas to H3's 32-pixel grid, builds the packed H3 audio/video latent containing the requested temporal packet, encodes any source or reference images through the VAE, and returns positive conditioning ready for the sampler.

One subtlety worth knowing: if you set image_to_image (FL2VA) with a one-frame profile, the node silently switches to REF2VA-style source conditioning. That's not a bug - FL2VA's frame-0 anchor would occupy the only output frame, which is impossible in a one-frame edit. Multi-frame image editing keeps the original FL2VA keyframe path.

Inputs that matter

The inputs, cut down to the ones you'll actually touch:

• clip - the MiniMax H3 Qwen text/vision encoder (also used for reference images, which is why REF2VA needs it).
• mode, prompt, width/height - the basics.
• frame_preset - single image | 1 frame, recommended | 5 frames, or 9/13/20. The one-frame profile is meant for REF2VA and the experimental H3 image VAE; multi-frame profiles use the standard video VAE.
• preserve_strength - prompt-language preservation for edits. It is not denoise strength and doesn't touch the sampler schedule; a trap a lot of people fall into.
• source_fit - crop_center, contain_pad, or stretch for fitting source/reference content to the canvas.
• Optional: vae (required for I2I and REF2VA since images must be encoded - T2I ignores it), source_image, and reference_image_2 through reference_image_9 (REF2VA only; they're politely ignored in the other modes, and run_info tells you when).

Outputs and two gotchas

The outputs are the same shape as the dedicated prepare nodes: positive goes to your sampler's positive input, h3_latent goes to the sampler and then to Exact Frame Decode, fitted_source is a preview of the source on the generation canvas, requested_frames feeds the decoder, and optimized_prompt + run_info show you exactly what the node did (including warnings like "source_image is connected but ignored in Text to Image mode").

Two things to remember. Each reference socket represents exactly one picture - if an upstream node sends an IMAGE batch, only its first image is used, so your <Picture N> numbering stays stable. And in REF2VA you must state the role of every connected picture explicitly in the prompt; the node optimizes wording, it can't guess roles.

Install

Install is the pack, nothing more exotic: 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. Requires ComfyUI 0.30.0+ and the H3 model stack from the official ComfyUI MiniMax H3 guide - the pruned INT8 ConvRot diffusion model, Qwen text encoder, and the video VAE (the audio VAE is not needed; the latent carries audio metadata but this pack never decodes it to sound).