Zhenzhen's AI Radar · MiniMax H3 · 2026-11-06 Evening Report
This issue features 8 new versions: Strongly Recommended 0 · Recommended 2 · Worth Watching 2 · Listed Only 4
Editor's Note This issue reviews 8 new H3 versions. First recommendation: GrayColor Retro Anime Style LoRA: Unifies text-to-video output into the look of 80s cel animation, with trigger words and strength provided, plus 4 visible example videos; the fisheye perspective beast-ear boy anime LoRA has detailed prompts, also rated A. tsolful's Turbo collection adds a Pika H3 Flex 8-step compatible version, and mo_akkakk's tiled + temporal window upscaling node workflow is suitable for video-to-video and secondary upscaling, but both lack visible examples, rated B. Clementine's H3 version only has Krea 2 parameters, AI Girl 40 filename says 39, Ghost Rider v1.1 this version only has JSON (weights uploaded in v1.0 entry), all downgraded to C.
Recommended (2)
[A] GrayColor Retro Anime Style · epoch7
LoRA · New Model · Author GrayColor · Released 11-06 20:25 · Downloads 12
80s retro anime style LoRA: Unifies H3 text-to-video visuals into the look of old-school cel TV animation, with a set of trigger words registered by the author.
Images
Reasons for Recommendation
- New model; serves 2D retro anime shorts, MVs, and nostalgic-style text-to-video, can be combined with character descriptions
- Trigger words and strength are provided: gc_anime, 1980's anime style, retro anime, strength approx. 0.8
- Includes 4 visible example videos, allowing you to check the style before deciding to download; less than 1 hours since release with 12 downloads and 3 likes (only indicates attention level)
How to Use
- Load gc_anim_minimaxh3_3432945_epoch_7.safetensors, strength approx. 0.8, start prompt with gc_anime, 1980's anime style, retro anime,
- Steps, CFG, and sampler are not provided by the author, start with your usual H3 text-to-video settings
Risks and Limitations
- Description is only two lines, no parameters other than strength; additionally, 4 example segments are hidden due to rating
- License only allows on-site rental, commercial image output is not allowed; this entry was not tested
Trigger Words: gc_anime, 1980's anime style, retro anime,
[A] twink chibi video 1 · v1
LoRA · New Model · Author geminigooner · Released 11-06 14:59 · Downloads 25
2D anime video LoRA: Extreme fisheye perspective, strong foreshortening, and anime opening-style camera rush actions, with themes centered around beast-ear boys and neon cyber color schemes.
Images
Reasons for Recommendation
- New model; serves high-energy anime shots in text-to-video such as "character lunges toward camera, points at viewer," author claims 2D lines do not blur during camera rushes
- Prompt writing is detailed: style and perspective words, action and camera movement words, and character appearance words are listed in three groups, with 3 full prompts for preview
- Includes 1 visible example videos; free, license allows commercial image output and sales
How to Use
- No trigger words registered; start with the author's style words: 2D anime style, high-energy animation, extreme fish-eye perspective, foreshortening, dynamic angle
- Write actions and camera movements like lunges forward toward the camera, pointing right at the viewer, fast-paced camera zoom, etc.; strength and steps are not provided by the author
Risks and Limitations
- Author's description lists Epoch 1–4 and recommends Epoch 2, but this model page currently only has one v1 version, unclear which Epoch it corresponds to
- Only 1 visible examples, another 2 segments are hidden due to rating; themes are concentrated on beast-ear boys, effect on other characters is unknown; this entry was not tested
Worth Watching (2)
[B] Minimax H3 Turbo LoRas · pika-H3-flex-Turbo-8steps
LoRA · Version Update · Author tsolful · Released 11-06 18:52 · Downloads 51
New member of the H3 acceleration LoRA collection: A compatible version of Pika Labs' Pika H3 Flex 8-step acceleration LoRA, allowing it to be used with ComfyUI's built-in LoRA loader and the streamlined H3 base model.
Images
Reasons for Recommendation
- Version update: The 11th version of the collection, adds pika_h3_flex_fl2va_v1_bf16; author states they did not train the original weights, only made a compatible version, and credited original authors item by item
- Serves users who want to reduce steps for H3 text-to-video (FL2VA); author's current suggestion is 6–8 steps, marked as Work In Progress (WIP)
- Entire collection has accumulated 54262 downloads and 113 comments, this version has 51 downloads in 2 hours (only indicates attention level)
How to Use
- Use ComfyUI's built-in LoRA loader to attach to the H3 base model, strength 1.0, steps 6–8; sampler, scheduler, and shift are not separately provided by the author for this version, can refer to er_sde + simple settings of other acceleration LoRAs on the same page
Risks and Limitations
- Parameters for this version are marked as WIP by the author; no visible examples (2 hidden due to rating), version description section is empty
- Original weights come from Pika Labs, confirm original license before use; this entry was not tested
- Card image is an AI-generated diagram, not the actual model effect; author has no publicly available general-rated examples
[B] Minimax-H3_latent_upscaler · v1.0
Workflow · New Model · Author mo_akkakk · Released 11-06 20:24 · Downloads 18
Video-to-video / upscaling workflow paired with the author's custom ComfyUI nodes: Splits high-resolution long videos into spatial tiles and temporal windows for batch processing, author claims motion transfer without drift or ghosting.
Images
Reasons for Recommendation
- New model; serves two scenarios: performing video-to-video at high resolution in the first stage, or performing second-stage upscaling on already generated videos
- Author claims to have run 2.5 MP on an RTX TI GPU with 12 GB VRAM; nodes are published on Comfy Registry (comfyui-minimax-h3-temporal-window-spatial-tiles)
- Author explicitly states it is not recommended for generating from scratch: generate the first stage normally, then use this for upscaling
How to Use
- Install the author's nodes from Comfy Registry, import Minimax-H3_latent_upscaler.json; parameters such as tile size, window length, and denoising strength are not listed in the description, need to check settings within the workflow
Risks and Limitations
- No visible examples (2 hidden due to rating), effect is only described in text by the author
- Depends on the author's custom nodes and H3 model files, specific list needs to be verified by opening the workflow; this entry was not tested
- Card image is an AI-generated diagram, not the actual model effect; author has no publicly available general-rated examples
Listed Only (4)
- [Krea 2 DiT] Clementine (zclclementine) — Fictional AI Influencer & Consistent Character · v1.0 - MiniMax H3 Charact — LoRA: H3 character LoRA for fictional influencer Clementine, ported from Krea 2 version, trigger word zclclementine.
- AI Girl: Fictional Women Series40 / Fictional Women Series40 MiniMax H3 · v1.0 — LoRA: H3 character LoRA for the 40th entry of the "Fictional Women Series," generating the same fictional Japanese woman.
- MiniMax H3 Video to Video Workflow · v3.0 — Workflow: Same author's H3 video-to-video workflow v3.0, modified to work with the aforementioned tiled + temporal window nodes for single-stage video editing.
- KinetiX Ghost Rider Hellfire Wheels & Molten Ground Concept Slider · v1.1 — LoRA: Ghost Rider-style hellfire wheels and molten track effect slider v1.1; this version only uploaded a JSON, new weights are placed by the author in the v1.0 entry.
Data: Civitai Public API (Base Model MiniMax H3), inclusion window 11-06 08:31 → 11-06 21:01 (Beijing Time). Recommendation reasons reviewed by editors; download counts only indicate attention level.