Zhenzhen's AI Radar · MiniMax H3 · 2026-11-08 Evening Report
This issue features 8 new versions: Strongly Recommended 0 · Recommended 4 · Worth Watching 3 · Listed Only 1
Editor's Note This issue reviewed 8 new H3 versions: 4 rated A, 3 rated B, 1 rated C, and no S. The top recommendation is the CROWN VFX & High-Speed Combat Boost LoRA: it trains only layers related to video motion and effects, enhancing martial arts, weapon actions, and fluid/lightning effects, with two sets of sampling parameters provided. Also included are a Pika person-swap workflow that replaces the person in a video with a single photo (preserving original audio), a P-H3-Image workflow for fixed-face random outfit changes, and two versions of the realism slider ported to the HyperFlow distilled model.
Recommended (4)
[A] CROWN VFX & High-Speed Combat Boost | MiniMax H3 · v2.0 - VideoOnly r64
LoRA · New Model · Author: NineCrowns · Published 11-08 13:00 · Downloads 117
A self-trained H3 LoRA for enhancing VFX and high-speed combat: it targets only layers related to video motion and effects, boosting fast-paced martial arts and weapon actions, fluid/plasma trails, and lightning effects, while preserving facial features when generating video from reference images.
Images
Reasons for Recommendation
- New model; serves action scenes and xianxia/sci-fi combat videos, usable for text-to-video, image-to-video, and reference-image-to-video
- Comprehensive author parameters: Rank 64, with two sampling settings provided (24 steps ManualSigmas + euler, or 20 steps euler + beta), and the strength of the accompanying LoRA is specified as 0.45
- License allows commercial use and sales; approximately 117 downloads in about half a day since release (indicating attention level only)
How to Use
- Include the trigger word sd25_vfx_combat in the prompt; use the author's 20 step euler/beta or 24 step euler settings, combined with MiniMaxH3SigmaShift
Risks and Limitations
- "100% face preservation" is the author's claim; only 1 video examples are provided
- This entry has not been tested
Trigger Word: sd25_vfx_combat
[A] p-h3-image - Random Pose & Outfit with Fixed Face Workflow · v1.0
Workflow · New Model · Author: TEKUA · Published 11-08 17:55 · Downloads 49
A "fixed face, random pose and outfit" workflow for P-H3-Image: three reference images provide pose, clothing, and face/hair respectively, with the workflow automatically rotating through a folder of poses and outfits on each run.
Images
Reasons for Recommendation
- New model (workflow); serves batch generation of outfit and pose variations for the same character
- The face reference image remains fixed, while poses and outfits rotate automatically from two folders; documentation is primarily in Japanese but steps are clear
- 5 visible examples; license allows commercial use and sales
How to Use
- Import minimax_p_h3_random_image_v1.json, place face/hair images in Reference 3, specify pose and outfit folders separately, and run repeatedly; requires the image loading node from ComfyUI-AITECCAFE-Toolkit
Risks and Limitations
- Depends on third-party node packages; parameters are not individually listed by the author
- This entry has not been tested
[A] MiniMax H3 Pika Person Swap | One Photo, Keeps the Sound · v1.0
Workflow · New Model · Author: AusBoss · Published 11-08 10:10 · Downloads 221
An H3 workflow that replaces the person in a video with another person using a single photo: scene, action, camera movement, and original audio are all preserved.
Approximately 221 downloads in about half a day since release, making it the most attention-grabbing entry this issue (indicating attention level only).
Images
Reasons for Recommendation
- New model (workflow); serves short video person swaps and character replacements
- Author specifies testing on ComfyUI 0.38.0 + AusBoss nodes 2.21.2, with 24, 16, 12, and 8 GB VRAM; clips are cropped to a single shot, max 10 seconds
- Prompt writing guide provided: first state who to replace, then describe the new person's hairstyle and face on a new line
How to Use
- Load and crop the clip in Video - Your clip, place the new person's photo in Picture 1, write the prompt as "Replace the man in the red shirt with @1." and describe @1's appearance on the second line; select High VRAM if you have 24 GB or more
Risks and Limitations
- Person swapping involves portrait rights; use only with authorized photos and videos
- 1 examples are hidden due to rating; this entry has not been tested
[A] Modern Beauty - MMH3 · v2.0
LoRA · Version Update · Author: DuckyDuo · Published 11-08 10:38 · Downloads 232
v2 of the "Modern Beauty" style LoRA: adjusts H3 portrait aesthetics to be more modern and adds a trigger word.
Images
Reasons for Recommendation
- Version update: update notes state a larger and better dataset, a better tagging model, 30% more training steps, and a new trigger word MdrnBty
- The v1 on the same page accumulated 602 downloads over two issues after release (indicating attention level only)
How to Use
- Write MdrnBty in the prompt, strength 1.0
Risks and Limitations
- Only 1 video examples, no comparison with v1
- License allows on-site rental only; this entry has not been tested
Trigger Word: MdrnBty
Worth Watching (3)
[B] MMH3 Realism Slider - HyperFlow H3 Flash port · ref2va
LoRA · New Model · Author: voidhunter77 · Published 11-08 18:03 · Downloads 92
A LoRA porting @alcaitiff's H3 realism slider to the distilled HyperFlow H3 Flash (reference image/first-last frame model ref2va), with unchanged weights but renumbered layers.
Images
Reasons for Recommendation
- New model; solves a specific problem: HyperFlow removed entire layers from H3, causing the original LoRA to load onto incorrect layers; ref2va removed layer 9, and this version renumbers so all 100 weights load correctly
- Author's test conditions: HyperFlow 6 steps, 768x1344; recommended strength 1.0–1.5; link to original provided
How to Use
- Use with minimax_h3_ref2va_hyperflow_flash_14b, strength 1.0–1.5
Risks and Limitations
- Only applicable to the ref2va version of HyperFlow; use the original slider for the original H3
- 2 examples are hidden due to rating; this entry has not been tested
- The card image is an AI-generated diagram, not the model's actual output; the author has no publicly available safe-for-work examples
[B] MMH3 Realism Slider - HyperFlow H3 Flash port · fl2va
LoRA · Version Update · Author: voidhunter77 · Published 11-08 18:03 · Downloads 50
The fl2va version of the same realism slider port (for the HyperFlow model generating video from text prompts only), renumbered after removing layers 6, 21, and 22.
Images
Reasons for Recommendation
- Version update: on the same page as the ref2va version; author states 98 of the 100 weights load, and ComfyUI warnings about one layer not loading are normal
- Recommended strength 1.0–1.5, test conditions same as the ref2va version
How to Use
- Use with minimax_h3_fl2va_hyperflow_flash_14b, strength 1.0–1.5
Risks and Limitations
- Only applicable to the fl2va version of HyperFlow; 4 examples are hidden due to rating
- This entry has not been tested
- The card image is an AI-generated diagram, not the model's actual output; the author has no publicly available safe-for-work examples
[B] Progressive Acceleration MMH3 Slider: Smoothly accelerates the speed of any action or motion over time. · KinetiX Acceleration V2.0
LoRA · Version Update · Author: make676767 · Published 11-08 12:33 · Downloads 14 · Paid
Paid "Progressive Acceleration" action slider V2: makes actions in the video speed up over time.
Images
Reasons for Recommendation
- Version update: update notes describe this as a concept slider controlling the speed curve of actions between frames
- Author provides recommended strength 4, usable range -8 to 8, and an example workflow
How to Use
- No trigger word needed, strength 4 (adjustable between -8 and 8, negative values likely slow down, author did not elaborate)
Risks and Limitations
- Paid download; all 5 examples are hidden due to rating
- This entry has not been tested
- The card image is an AI-generated diagram, not the model's actual output; the author has no publicly available safe-for-work examples
Listed Only (1)
- [Krea 2 DiT] Clementine (zclclementine) — Fictional AI Influencer & Consistent Character · v1.0 - MiniMax H3 RefLoRA — LoRA: H3 "RefLoRA" for the fictional influencer character Clementine: packages character LoRA weights and RefMod reference latents into a single file.
Data: Civitai public API (base model MiniMax H3), inclusion window 11-08 09:33 → 11-08 21:01 (Beijing Time). Recommendations have been reviewed by editors; download counts reflect only attention levels.