Image Generation

A 66 GB BF16 trunk, one single-head node and a 172 MB LoRA: how T8 pushed MiniMax H3 Image from 6 steps to 50

November 22, 20260 views阅读中文原文
A 66 GB BF16 trunk, one single-head node and a 172 MB LoRA: how T8 pushed MiniMax H3 Image from 6 steps to 50

TL;DR

MiniMax H3 Image is a roughly 33.1B-parameter image model that shipped as a six-step, CFG-1 distillation, with a single checkpoint covering both text-to-image and prompt-driven editing. On November 21, the Chinese creator T8 (T8star-Aix) published an unofficial ComfyUI package: a full BF16 trunk converted to ComfyUI's layout, a custom "single head, multi-step" node, and a separately trained MLP100 LoRA. The goal is to run 50 steps at CFG 3 and get more natural materials than the native six-step path produces.

T8 is explicit about the boundaries. This is not distillation removal and not a "99% teacher recovery." It is an approximation: the six released PDD output heads are averaged into one head within intervals, and a LoRA compensates for the material degradation that shows up at higher step counts. Quality was judged by hand on ten fixed samples, of which eight passed and two failed outright.

Why the model "can only run six steps"

The released H3 Image weights use a multi-head PDD structure: six output heads, where each sampling step is handled by one specific head. That is why the official recipe is six steps, CFG 1 and no negative prompt. Run it the way you would run an ordinary diffusion model, at 50 steps, and the model predicts with a head that never learned those steps — faces break into blocks.

T8 demonstrates the comparison on video. Native six-step output is usable but low in texture, and only partly rescued by rendering at 16 MP or stacking quality LoRAs. Force the native weights to 50 steps and the image falls apart.

He tried two correction routes. The first subtracted an official PDD delta, but the PDD that Alibaba published earlier is not the same one, and H3 Image was derived from a video model with the audio branch cut away, so the subtraction did not land correctly. The second route is what shipped: instead of subtracting from the original, build a node that collapses the six heads into one by weighted average, then train a LoRA to compensate for what multi-step sampling loses.

The three pieces

  • Full trunk (minimax_h3_image_full_bf16_comfy_qkvfix-T8.safetensors): 66.29 GB, the official shards renamed and rearranged into ComfyUI's QKV layout, keeping the full AdaLN and the original release precision. It is not an INT8 dequantization and not a pruned model. The PDD six-head structure is still inside, so the native six-step route keeps working.
  • Single-head node (H3 Image Single Head · T8): averages the six heads into one output head within intervals, editing only an in-memory clone rather than overwriting the trunk file. The node repository carries a mean_head.safetensors asset that must be kept alongside it.
  • direct MLP100 LoRA (experimental_direct_mlp100_rank16-T8.safetensors): 172.07 MB at rank 16 / alpha 16. The trunk is frozen; the adapter was trained for 224 updates across 100 MLP modules in 50 layers, with no attention adapters. It is a newly trained approximation, not a negative subtraction of the original speed LoRA.

The text encoder and VAE reuse existing community files: an INT8 Qwen3-VL-32B text encoder (27.14 GB) and an FP16 image VAE (1.17 GB). So the pipeline is not BF16 end to end — only the trunk is BF16 with FP32 mixed storage.

How to use it: wiring order and three settings you should not change

  1. Put the trunk in ComfyUI/models/diffusion_models/, the LoRA in loras/, and the text encoder and VAE in their folders.
  2. Install the node with git clone https://github.com/T8mars/Comfyui-h3-image-single-head-T8.git custom_nodes/Comfyui-h3-image-single-head-T8, then restart ComfyUI. A node mirror also exists in the Hugging Face repo; install one or the other, never both.
  3. Import the plain .json workflows under workflows/direct_mlp/. The .api.json files are submission payloads, not UI workflows.

The graph is a fixed four-stage chain:

```text

UNETLoader → H3ImageSingleHead → LoraLoaderBypassModelOnly (+1) → CFGGuider (3)

```

Three common mistakes:

  • The single-head node must sit directly on a trunk that no LoRA has touched. In the wrong order the weighted average is meaningless.
  • The LoRA must go through the bypass loader at strength +1, not -1. The bypass loader computes the LoRA as a separate layer, recomputes the corresponding trunk block, then merges. A normal LoRA merge loader produces visibly different images.
  • Use Euler, 50 steps, CFG 3 and shift 3, with the 51-point sigma table embedded in the workflow, and do not stack another SHIFT node on top. T8 exposes the sigma table separately so it can be edited.

Results: ten samples, eight passes, two failures

The comparison holds the trunk, the mean head, the prompt, seed, resolution, sampler and sigma grid fixed, and changes only whether direct MLP100 is loaded. The left column is the 50-step control without the LoRA; the right column has it.

In the eight passing samples the change is mostly material. Blocky hard edges on an elderly man's skin decrease while age lines survive. Linen goes from a fractured relief look to continuous weave. Fur stops looking carved. Watercolor loses some of its paper embossing while text and style remain. The glass and two-handled bowl keep their material separation and all three joints.

The two failures are specific. In the umbrella sample the red canopy looks more natural, but the full shaft and handle are still missing. In the wall-sign edit the large characters are mostly removed, yet a faint trace remains at the top and the figure plus non-target details drift.

That places the package as "let me get more natural materials at multi-step," not "let me edit precisely." T8 also notes that 8/10 is a joint human check on that fixed sample set by him and his team — not teacher similarity, not a large blind test — and that generalization to other subjects and seeds is unproven.

Who it is not for

  • Anyone hoping for the original teacher model. The repository contains no matching original PDD trunk delta and no unknown teacher weights to compare against, so no restoration claim can be made.
  • Machines short on VRAM and RAM. The test rig was an RTX 4060 Ti with 16 GB of VRAM and 128 GB of system RAM, and the 66 GB of weights need heavy system memory and offloading. Running on 16 GB of VRAM does not mean a 12 GB card or a low-memory box will manage, and it says nothing about speed.
  • Commercial users. The H3 Image community license excludes the EU, UK, South Korea and the United States, and has separate authorization requirements for hosted services and revenue thresholds. Open download does not change the upstream terms.
  • Anyone doing precise local edits. The failure samples show non-target preservation is still unstable.

Cost and alternatives

The real cost is download time: a 66.29 GB trunk, a 27.14 GB text encoder, a 1.17 GB VAE and the 172 MB LoRA. If speed is all you want, the official six-step route needs neither the node nor MLP100 — workflows/native_pdd6 is the native validation workflow, running BasicGuider as the CFG-1 equivalent.

Other acceleration options exist in the community, such as a two-step pDMD LoRA whose merged checkpoint is around 40 GB and whose model card claims about 20 seconds for a 2048 px image on an H200. That points in the opposite direction — fewer steps rather than more steps for better material — so do not read the two as the same thing.

Next step

If you already run the H3 Image six-step workflow, try it in this order: swap in only the BF16 trunk first and confirm native six-step output matches what you had; then add the single-head node and MLP100, moving CFG from 1 to 3 and steps to 50; finally rerun with your own subjects and seeds instead of trusting the author's ten fixed samples. T8 says the training walkthrough will be a separate episode; this one covers weights and usage only.

Sources

  • Bilibili video: [T8开源最强 Minimax H3 Image 混合精度 QKV BF16 模型 + 100MLP 微调 Lora + 单头注意力节点,50步 CFG3 效果对比测评](https://www.bilibili.com/video/BV1nbyez4EsR) (T8star-Aix, November 21, 2026)
  • Model card and weights: [t8star/Minimax-H3-Image-Exp-Comfy](https://huggingface.co/t8star/Minimax-H3-Image-Exp-Comfy) (updated November 21, 2026)
  • Single-head node: [T8mars/Comfyui-h3-image-single-head-T8](https://github.com/T8mars/Comfyui-h3-image-single-head-T8)
  • Base model specs and native usage: [ComfyUI docs · MiniMax-H3-Image](https://docs.comfy.org/tutorials/image/minimax/minimax-h3-image)
  • Base weights: [MiniMaxAI/MiniMax-H3-Image](https://huggingface.co/MiniMaxAI/MiniMax-H3-Image)

This article is compiled from the creator's original video transcript, the model card and ComfyUI's official documentation. All performance judgments come from the author's own fixed-sample comparison and are not an independent third-party evaluation.

Sources

Comments

0 published
Log inorsign up to leave a comment.

Loading comments...