Zhenzhen's AI Radar · MiniMax H3 Image · 2026-11-19 Evening Report
This issue features 5 new versions: Strongly Recommended 0 · Recommended 2 · Worth Watching 3 · Included Only 0
Editor's Note This issue reviews 5 new versions of H3 Image, with 2 rated A and 3 rated B. The top recommendation is the Exact-adaLN NVFP4 quantized base model: 12.61 GB. The author claims that RTX 50 series graphics cards with 16 GB of VRAM can also generate 4 MP images. The modulation part uses exact lookup tables, resulting in zero error compared to the official version. Installation and sampling settings are fully documented. The author provided strength settings for a graphic novel style LoRA trained on their own 100 original works, rated A. The ported versions of ArsMachina's three classic styles (80s fantasy, 60s psychedelic, and religious illuminated manuscripts) are all paid downloads, rated B.
Recommended (2)
[A] MiniMax-H3-Image Exact-adaLN NVFP4 · v1.0
Base Model / Checkpoint · New Model · Author lepetitprince716355 · Released 11-19 13:18 · Downloads 19
NVFP4 quantized base model for MiniMax H3 Image (12.61 GB): It compresses the reprojection layer to 4 bits while replacing the common pruned version's approximation of adaLN modulation with exact lookup tables, allowing RTX 50 series graphics cards with 16 GB of VRAM to generate 4 MP images.
Suitable for users with 50 series cards having 16 GB of VRAM; if you have a card with 24 GB or more, running the BF16 version does not pose VRAM pressure, so you may not necessarily need this one.
Images

Reasons for Recommendation
- The changes are very specific: The 6-step PDD sampling of H3 Image only calls the model at 7 fixed time points. Based on this, the author implemented adaLN as a 7-point lookup table. The author claims zero error compared to official modulation 0 ULP; common pruned versions use 8-dimensional curve approximation, which introduces errors.
- Size is 12.61 GB. The author claims it is 3.2 times smaller than the pruned BF16 version; resident VRAM is 11.75 GiB, peak VRAM for 4 MP text-to-image is 12.34 GiB, and peak VRAM for dual-reference image editing is 14.00 GiB. With text encoder offloading, it fits on 16 GB cards.
- The author provided speed metrics: 4 MP, 6 steps take approximately 10.0–10.5 seconds (about 1.65 seconds per step). The author claims it is 1.42 times faster than the BF16 version with exact adaLN; 7 example images are visible, and HF mirrors, checksums, and build scripts are all public.
How to Use
- Replace the H3 Image base model loading; use CLIPLoader (type: minimax) to load the BF16 or NVFP4 AWQ version of qwen3vl_32b_minimax_h3 for the text encoder, and use minimax_h3_image_vae for the VAE.
- For sampling, use SamplerCustomAdvanced + euler + BasicGuider + ManualSigmas (6-step PDD, no CFG), fill in the 7 sigma values provided by the author. The author claims this is closer to the official schedule than KSampler simple.
- Drag any example PNG from the model page into ComfyUI to load the corresponding workflow.
Risks and Limitations
- 4 bit acceleration only works on Blackwell graphics cards such as the RTX 50 series; older cards will dequantize on the fly. It requires newer versions of ComfyUI, comfy-kitchen, and PyTorch with CUDA 13.
- This is an unofficial version; the upstream MiniMax-H3-Image community license excludes regions such as the EU, UK, South Korea, and the US.
- Accuracy and speed figures are as claimed by the author; it was just released and has very few downloads.
[A] Suideo's Graphic Novel Style for H3-Image · v1.0
LoRA · New Model · Author redhotcherrypop · Released 11-19 18:05 · Downloads 18
A graphic novel illustration style LoRA, natively trained for H3 Image. The author used 100 works manually rendered over many years as the training set.
Images

Reasons for Recommendation
- New style: The author emphasizes that it was trained for H3-Image itself, not converted from a video model; the training images are the author's original works with clear provenance.
- The author provided a strength range: 1.2–1.6 when running a 6-step workflow without stacking other LoRAs; no trigger word is needed, add kafkaeffect4 if the style is not obvious.
- Free, license allows selling.
How to Use
- Load the LoRA with strength 1.2–1.6 (for 6-step workflows without stacking other LoRAs); adjust manually if stacking LoRAs or using different step counts.
- No trigger word needed; add kafkaeffect4 if the style does not appear.
Risks and Limitations
- Only 2 example images.
- Single epoch version; the author did not specify which subjects it is suitable for.
Worth Watching (3)
[B] 80s Fantasy Movie · MiniMax H3 Image
LoRA · Ported · Author ArsMachina · Released 11-19 18:34 · Downloads 4 · Paid
H3 Image version of the LoRA for early 80s fantasy movie still style: castles, monsters, and fantasy characters with the film grain texture of old movies.
Images

Reasons for Recommendation
- Ported: The same page already has over ten base model versions including Krea 2, MOSS, Flux.3 Klein, Qwen, and SDXL; the entire model page has 46725 downloads. The author explains that the dataset was redone in the Midjourney 7 era, but differences for the H3 Image version are not specified.
- Serves fantasy posters and cinematic concept art; trigger words are provided, and 9 examples are visible.
How to Use
- Write ArsMovieStill, 80s Fantasy Movie Still in the prompt.
- Strength and step count are not provided by the author (the author's "recommended settings" are a joke).
Risks and Limitations
- Paid download.
- Training images are from Midjourney outputs; the license does not allow secondary modification and publication.
- Another 6 sensitive-level examples are hidden.
Trigger Words: ArsMovieStill, 80s Fantasy Movie Still
[B] 60s Psychedelic Movie · MiniMax H3 Image
LoRA · Ported · Author ArsMachina · Released 11-19 19:10 · Downloads 3 · Paid
H3 Image version of the LoRA for 60s psychedelic movie still style: high-saturation colors and psychedelic light effects of old movie scenes.
Images

Reasons for Recommendation
- Ported: The same page already has versions for MOSS, Krea 2, Flux.2 Klein, SDXL, etc.; the entire model page has 13270 downloads. Differences for the H3 Image version are not specified.
- Trigger words are provided in the description, and 9 examples are visible.
How to Use
- Write ArsMovieStill, movie still from a 60s psychedelic movie in the prompt.
- Strength and step count are not provided by the author.
Risks and Limitations
- Paid download.
- License does not allow secondary modification and publication.
- Another 4 sensitive-level examples are hidden.
[B] Religious Illumination Fantasy · MiniMax H3 Image
LoRA · Ported · Author ArsMachina · Released 11-19 18:08 · Downloads 0 · Paid
Medieval manuscript illumination and religious icon painting style, used for fantasy themes: hero portraits, mythical beasts, star charts, and epic scenes.
Images

Reasons for Recommendation
- Ported: The same page already has versions for MOSS, Krea 2, Anima, Flux.2 Klein 9B, and ZImageBase; the author says this is the first time training for H3 Image, using 3k steps.
- Keywords are provided, and 9 examples are visible.
How to Use
- Write Religious Illumination Fantasy in the prompt.
- Strength and step count are not provided by the author.
Risks and Limitations
- Paid download, license does not allow secondary modification and publication.
- Another 1 sensitive-level examples are hidden.
Data: Civitai Public API (base model MiniMax H3), inclusion window 11-19 08:31 → 11-19 21:01 (Beijing Time). Recommendation reasons have been reviewed by editors; download counts only reflect attention.