Minimax H3 Image INT8 / INT6 / INT4 ConvRotRecommended settings: 6 steps, CFG 1, PDD sampling (official). My settings LCM/simple 6stepsQuantized weights of the Minimax H3 Image model, packaged for ComfyUI. Four precision tiers are available so you can pick the right point on the speed/quality / VRAM/RAM curve for your hardware.About these variants Four precision levels are available (all sizes are pruned checkpoints), trading VRAM and speed against generation quality:<ul><li>BF16: full-precision reference (40.2 GB). Highest quality, largest footprint. Use it if you have the VRAM and want the best result to compare the others against.</li><li>INT8 ConvRot (W8A8): 8-bit weights and 8-bit activations (19.54 GB). Near-lossless quality, runs on INT8 tensor cores (RTX 30-series and up). This is the default recommendation.</li><li>INT6 (W6A8): 6-bit weights with 8-bit activations (15.61 GB). A middle ground between INT8 and INT4: noticeably smaller than INT8 with a small quality trade-off. A good pick when INT8 is close to fitting but doesn’t quite. Uses the W6A8 kernel path from <a target="_blank" rel="ugc" href="https://github.com/Comfy-Org/comfy-kitchen/pull/191">comfy-kitchen PR #191</a>.</li><li>INT4 (W4A8 mixed): 4-bit weights with 8-bit activations (12.77 GB). Smallest footprint of the group. Uses a ConvRot codebook layout (<a target="_blank" rel="ugc" href="https://github.com/Comfy-Org/comfy-kitchen/pull/90">comfy-kitchen #90</a>) that decodes to INT8 at compute time, so it runs on the same INT8 hardware as W8A8 with the same throughput. The gain is memory, not speed. Larger quality trade-off than INT6.</li></ul>All quantized variants keep the sensitive layers (I/O, modulation, embeddings, output head) in BF16. If you’re not sure, start with INT8. Move to INT6 if you need a bit more room, and to INT4 if you’re actively hitting VRAM limits. The quality gaps from INT8 to INT6 and from INT6 to INT4 are small enough that both are good trade-offs on tight budgets.About the model Minimax H3 Image is MiniMax’s open-weight image model, built on H3: it uses the 33.1B H3 transformer with the audio branch removed. One checkpoint handles both:<ul><li>t2i (text to image): run with no images connected. Default output is 2048×2048.</li><li>Image editing / reference generation: connect images starting at <code>image_1</code> (the image being edited); the rest are references. Refer to them in the prompt as <code>&lt;Picture 1&gt;</code>, <code>&lt;Picture 2&gt;</code> and so on. Supports up to 16 reference images. Connect the VAE when editing so references are anchored as latents.</li></ul>Highlights<ul><li>Native output up to 4096×4096 (16 MP), generated at that size rather than upscaled. Native presets: 1:1 4096×4096 · 4:3 4736×3552 · 3:2 5024×3360 · 16:9 5472×3072 · 21:9 6272×2688 (4:3, 3:2 and 16:9 also work in portrait).</li><li>Step- and CFG-distilled: 6 steps at CFG 1. The model uses no negative prompt.</li><li>Long prompts work best: up to 4096 tokens. Full, detailed descriptions give the best results. <a target="_blank" rel="ugc" href="https://huggingface.co/MiniMaxAI/MiniMax-H3-Image-PE">MiniMax-H3-Image-PE</a> is the official prompt expander for t2i.</li><li>Not supported: masked edits, transparent backgrounds.</li></ul>Requirements<ul><li>ComfyUI at a version that supports Minimax H3 Image natively (use the “MiniMax H3 Image Conditioning” node, or load the MiniMax H3 Image template).</li><li>Text encoder: the same Qwen3-VL encoder as Minimax H3 video (e.g. <code>qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors</code>). If you already run H3 video, you have it.</li><li>VAE: <code>minimax_h3_image_vae_fp16.safetensors</code> from <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3-Image">Comfy-Org/MiniMax-H3-Image</a>. The video VAE won’t work.</li></ul>License: <a rel="ugc" href="https://huggingface.co/MiniMaxAI/MiniMax-H3-Image/blob/main/LICENSE">MiniMax-H3-Image Community License Agreement</a>.

Files with SHA256: 4647bdf553a32aa4ae55d9eb28322be42d32611d333c19a097d2845ae8a9b7a8