Original title: vyrekai/minimax-h3-image-comfyui-workflows
A ComfyUI workflow repository uploaded by vyrekai on 11 14, which does not include new weights. It provides three versions for MiniMax H3 Image: 8 GB, 12 GB, and 16 GB+. All can generate full-size images at 16 MP (e.g., 5472x3072). ComfyUI loads the model into the GPU in chunks; lower VRAM simply results in slower performance. The model files exclusively use official Comfy-Org int8 diffusion models, nvfp4 text encoders, and VAEs, requiring no custom nodes. A guide is also included to help you write prompts for H3 Image using chat models like ChatGPT and Claude.
Why it matters The official H3 Image model files are much larger than the VRAM of typical graphics cards. This workflow uses chunked loading to enable even 8 GB GPUs to generate full-size images.
Upvotes 62 · Likes 63 · Downloads 0
Reasons
- The author simulated three VRAM tiers on an RTX 5070 Ti with 24 GB of RAM: generating a 16 MP text-to-image takes about 66 seconds at 16 GB, about 63 seconds at 12 GB, and about 60 seconds at 8 GB. With the same seed, all three versions produce pixel-identical images; at 8 GB, ComfyUI's peak VRAM usage is 3.7 GB.
- The workflow includes aspect ratio selection, a toggle for 4 MP drafts vs. full size, 10 reference image slots (bypassed by default; enable with Ctrl+B when selected), and optional prompt expansion.
- The original files for the 8 example images are located in the repository's examples directory; the author states they have not been post-processed or upscaled.
Usage
- Select minimax_h3_image_8gb_vram.json, 12gb, or 16gb based on your VRAM and drag it into ComfyUI (author tested version 0.45.3). Download minimax_h3_image_pruned_int8_convrot (21 GB), qwen3vl_32b_minimax_h3_nvfp4_awq (15.7 GB), and minimax_h3_image_vae_fp16 (1.2 GB) from Comfy-Org/MiniMax-H3-Image, and place them in the corresponding directories under models.
- Sampling is pre-configured: SamplerLCM + simple, 6 steps, CFG 1.0, no negative prompt. Use 4 MP drafts to test prompts first, but note that switching to 16 MP with the same seed will produce a different image.
- Do not install the optional qwen3.5_9b prompt expansion model on 8 GB cards (it is disabled in the workflow). If VAE decoding causes VRAM overflow, reduce tile_size to 384 or 256.
Risks
- This is only a workflow and documentation, not a new model. The speed figures come from simulations where the author limited VRAM on an 5070 Ti; actual performance on real 8 GB / 12 GB cards (such as 4060, 3060) will be slower.
- Requires at least 24 GB of system RAM and approximately 38 GB of SSD space. On a 8 GB card, generating a 16 MP image using 4 reference images takes about 4 minutes.
- The workflow itself is licensed under Apache-2.0, while the model files are used according to the MiniMax-H3-Image community license. This entry has not been run.