What Is a MiniMax H3 LoRA?
A MiniMax H3 LoRA is a small adapter file loaded on top of the open H3-Base weights. It fine-tunes H3 output toward a specific style, character, camera behavior, motion vocabulary, or reduced step count. LoRAs are typically 100–500 MB and load in ComfyUI via a LoRA Loader node between the H3-Base checkpoint and the sampler.
H3 LoRAs are community-distributed. There is no single official LoRA catalog from MiniMax. This directory groups widely available LoRA archetypes by category and lists verified search sources for each.
How to Use a MiniMax H3 LoRA
- Download the open H3-Base weights (FL2VA and/or Ref2VA) from Hugging Face. See the MiniMax H3 Download guide.
- Verify the LoRA’s stated base model. FL2VA LoRAs and Ref2VA LoRAs are not interchangeable.
- In ComfyUI, connect the checkpoint into a LoRA Loader node and load the LoRA at the recommended strength.
- Match the sampler and step count recommended by the LoRA author. Turbo LoRAs need low steps; style LoRAs usually want standard steps.
- Test on 3–5 fixed seeds before committing to the LoRA for production work.
The full ComfyUI pipeline (H3-Base loader → LoRA loader → sampler → decoder) is documented in the MiniMax H3 ComfyUI Guide.
Turbo & Acceleration LoRAs
Step-distilled and framework-accelerated LoRAs that reduce H3 inference time.
H3 Turbo Acceleration LoRA
FL2VAA step-distilled LoRA that lets MiniMax H3-Base produce usable video in roughly 4–8 denoising steps instead of the standard 25–40. Trades some fine detail and prompt adherence for a large speedup on the same GPU.
Strength
0.8–1.0
Steps
4–8
Sampler
LCM / DPM++ / UniPC
CFG
3–5
Use cases
- Rapid prompt iteration
- Batch generation on a fixed GPU budget
- Low-VRAM setups with FP8 or GGUF weights
- Live storyboard demos
Cautions
- Do not run at 25+ steps — undoes the speed advantage and can degrade quality.
- Combined LoRA strengths above ~1.0 total often break motion coherence.
- Match the sampler the LoRA author recommends; step-distilled LoRAs are sampler-sensitive.
Where to find
H3 LightX2V-Compatible Acceleration LoRA
FL2VA or Ref2VAAcceleration LoRAs distributed for use with the open LightX2V framework, which applies step-distillation and runtime optimizations at the framework level. Stacks with FP8 / INT8 / GGUF quantization for further VRAM and throughput gains.
Strength
Framework-controlled
Steps
Framework-controlled
Sampler
Per LightX2V build
CFG
Per LightX2V build
Use cases
- Teams already running LightX2V for other diffusion video models
- Mid-range GPU inference with quantized weights
- Unified acceleration pipeline across multiple models
Cautions
- Requires a LightX2V build that officially supports MiniMax H3.
- Framework-level acceleration may not stack cleanly with every third-party ComfyUI node.
Where to find
Style LoRAs
Style-shifting LoRAs for cinematic, anime, or photoreal aesthetics.
H3 Cinematic Style LoRA
FL2VAStyle LoRAs trained to push MiniMax H3 output toward cinematic looks — anamorphic composition, filmic grain, controlled color palette, and shallow depth of field. Works best on text-to-video and first/last-frame image-to-video workflows.
Strength
0.5–0.9
Steps
25–35
Sampler
DPM++ 2M / UniPC
CFG
5–7
Common trigger words
Use cases
- Ad and commercial B-roll
- Trailer-style short clips
- Cinematic product presentations
- Music video atmosphere
Cautions
- Style LoRAs at strength > 1.0 can push contrast and grain past the point of realism.
- Some cinematic LoRAs bias the model toward night / low-light scenes; add explicit daylight cues when needed.
Where to find
H3 Anime & Manga Style LoRA
FL2VAStyle LoRAs that push MiniMax H3 toward 2D anime, cel-shaded, or manga-inspired aesthetics. Effective for short-form animated clips, opening credits, and stylized character shots.
Strength
0.7–1.0
Steps
25–35
Sampler
Euler A / DPM++ 2M
CFG
6–8
Common trigger words
Use cases
- Short-form anime clips
- Stylized character shots
- Animated opening / closing sequences
- Anime-style social content
Cautions
- Combining anime style with photoreal identity references usually produces uncanny hybrids.
- Long clips at strong anime strength may show frame-to-frame linework instability.
Where to find
H3 Realistic Portrait Style LoRA
FL2VAStyle LoRAs that improve photorealistic portrait rendering in MiniMax H3 — skin texture, catchlights, and micro-expression stability across frames. Best paired with a clear image reference for identity.
Strength
0.5–0.8
Steps
30–40
Sampler
DPM++ 2M / UniPC
CFG
5–7
Common trigger words
Use cases
- Portrait-forward short videos
- Talking-head style clips
- Realistic character close-ups
- Fashion and lifestyle content
Cautions
- Portrait LoRAs at high strength can over-smooth skin — reduce strength or add "natural skin texture" to the prompt.
- Do not stack multiple portrait LoRAs; effects rarely add linearly.
Where to find
Character LoRAs
Identity-locking LoRAs used with Ref2VA workflows to keep the same subject across frames.
H3 Character Consistency LoRA
Ref2VALoRAs trained to lock a specific character identity across frames when used with Ref2VA-family workflows. Combined with an image reference, they help hair, facial structure, and clothing stay stable across the whole clip.
Strength
0.6–0.9
Steps
25–35
Sampler
DPM++ 2M / UniPC
CFG
5–7
Use cases
- Recurring character in a series of clips
- Talking-head continuity
- Product mascot animation
- Multi-shot storytelling with the same subject
Cautions
- Requires a paired image reference for best results — the LoRA anchors identity but does not invent it.
- Do not combine two character LoRAs on the same subject — motion coherence usually breaks.
Where to find
Motion LoRAs
LoRAs that emphasize larger, more dynamic body motion.
H3 Motion Enhancement LoRA
FL2VA or Ref2VALoRAs that emphasize larger, more dynamic body motion and action in MiniMax H3 output. Useful when standard H3 renders read as too static for the target use case, such as sports, dance, or fight scenes.
Strength
0.5–0.8
Steps
30–40
Sampler
DPM++ 2M / UniPC
CFG
5–7
Common trigger words
Use cases
- Action and combat scenes
- Dance and choreography clips
- Sports highlight reels
- Dynamic camera-tracked motion
Cautions
- At high strength, motion LoRAs can distort limb geometry — reduce strength for tight close-ups.
- Pair with a motion or camera reference video for the most controllable output.
Where to find
Camera LoRAs
LoRAs that bias H3 toward specific camera behaviors and shot rhythm.
H3 Camera Behavior LoRA
FL2VA or Ref2VALoRAs that bias MiniMax H3 toward specific camera behaviors — orbits, dolly-ins, whip pans, handheld shake, or drone-style sweeps. Useful when the prompt alone does not reliably produce the intended shot rhythm.
Strength
0.4–0.7
Steps
25–35
Sampler
DPM++ 2M / UniPC
CFG
5–7
Common trigger words
Use cases
- Consistent orbit or tracking shots
- Handheld documentary look
- Drone-style aerial sweeps
- Push-in reveal shots
Cautions
- Camera LoRAs at strength > 0.8 can override subject placement — keep composition cues explicit in the prompt.
- Best combined with a video-to-video camera reference for the most predictable behavior.
Where to find
Quality & Detail LoRAs
LoRAs that improve fine detail, texture stability, and sharpness.
H3 Quality & Detail LoRA
FL2VA or Ref2VALoRAs that improve fine-detail rendering in MiniMax H3 — sharper textures, cleaner edges, and more stable micro-detail across frames. Useful when base H3 output reads as soft or washed out for close-ups and product shots.
Strength
0.5–0.9
Steps
30–40
Sampler
DPM++ 2M / UniPC
CFG
5–7
Common trigger words
Use cases
- Close-up product shots
- Textured fabric and skin detail
- Sharpening soft or low-light renders
- Final-pass enhancement before delivery
Cautions
- At strength above 1.0, detail LoRAs can introduce noise and oversharpening — verify on fixed seeds before production use.
- Detail LoRAs do not raise base resolution; pair with the 2K regenerate step for higher output size.
Where to find
VFX & Effects LoRAs
LoRAs specialized for particle effects, atmospheric FX, and stylized rendering.
H3 VFX & Effects LoRA
FL2VA or Ref2VALoRAs specialized for particle effects, atmospheric FX, and stylized rendering in MiniMax H3 — embers, sparks, dust, fog, rain, and magical glow. Helps prompts that call for layered effects render more reliably.
Strength
0.4–0.8
Steps
25–35
Sampler
DPM++ 2M / UniPC
CFG
5–7
Common trigger words
Use cases
- Ember, spark, and particle-heavy scenes
- Atmospheric fog, rain, and volumetric light
- Magic and energy effects for character shots
- Stylized transitions for social content
Cautions
- Heavy particle prompts at high strength can occlude the subject — keep the subject prompt explicit.
- VFX LoRAs bias output toward stylized rendering; reduce strength for photoreal scenes.
Where to find
Train Your Own MiniMax H3 LoRA
Training an H3 LoRA is realistic on a single 24 GB VRAM GPU for style / character tasks. Motion and camera LoRAs usually need higher-quality video datasets and more careful captioning than image-model LoRAs.
| Step | What to Do |
|---|---|
| Curate dataset | Collect 20–100 short clips (3–8 seconds each) that clearly demonstrate the target style, character, or motion. Consistent aspect ratio and lighting help. |
| Prepare captions | Write concise, structured captions per clip using the same phrasing patterns as the H3 native prompt guide. |
| Pick a target task family | Decide whether the LoRA is trained against H3-Base FL2VA (text / first-last frame) or Ref2VA (multimodal reference). This is not interchangeable. |
| Run training | Use a community H3 LoRA trainer that supports the chosen task family. Typical starting rank is 16–32; typical training length is a few thousand steps. |
| Evaluate | Generate 5–10 fixed test prompts across seeds and step counts. Compare motion coherence and identity stability against the untrained baseline. |
| Iterate | Adjust dataset caption quality, rank, and learning rate before re-training. Most usable H3 LoRAs require several iterations. |
MiniMax H3 LoRA FAQ
What is a MiniMax H3 LoRA?
A MiniMax H3 LoRA is a small adapter file loaded on top of the open H3-Base weights. It fine-tunes H3 output toward a specific style, character identity, camera behavior, motion vocabulary, or reduced step count without retraining the full base model.
Are MiniMax H3 LoRAs official?
The mainline open release is MiniMax H3-Base (FL2VA and Ref2VA task families). Community LoRAs are distributed by third parties on Hugging Face and Civitai. This page describes LoRA categories that widely exist in the community, not an official MiniMax product catalog.
Where can I download MiniMax H3 LoRAs?
Community LoRAs for H3 are distributed on Hugging Face and Civitai. Each LoRA category below links to a search on those platforms. Always verify the LoRA’s stated base model — an FL2VA LoRA and a Ref2VA LoRA are not interchangeable.
How do I load a MiniMax H3 LoRA in ComfyUI?
In ComfyUI, connect the H3-Base checkpoint into a LoRA Loader node, load the LoRA at the recommended strength, and feed the modified model into your sampler. The full pipeline (loader → sampler → decoder) is covered in the MiniMax H3 ComfyUI guide.
What is a Turbo LoRA?
A Turbo LoRA is a step-distilled LoRA that lets H3-Base produce usable video in far fewer denoising steps (typically 4–8 instead of 25–40). It is one of several ways to accelerate H3 inference. See the MiniMax H3 Turbo guide for the full breakdown.
Can I train my own MiniMax H3 LoRA?
Yes. Community H3 LoRA trainers exist for both FL2VA and Ref2VA task families. Training requires a curated dataset of 20–100 short clips with matched captions, and typically runs for a few thousand steps at rank 16–32.
How many H3 LoRAs can I stack at once?
Two or three LoRAs at moderate strength usually stack cleanly. Total combined strength above ~1.5 often breaks motion coherence. Do not stack two LoRAs that fine-tune the same axis — e.g., two character LoRAs on the same subject.
Do H3 LoRAs work with 2K output?
H3 LoRAs modify the H3-Base 768P generation stage. The 2K stage in the complete MiniMax H3 system is a separate H3-Regenerate-2K step. LoRAs applied at the base stage carry through to the 2K result.
What strength should I use for an H3 LoRA?
Most H3 LoRAs are stable in the 0.5–1.0 range. Turbo LoRAs typically want 0.8–1.0 with matching low step counts. Style LoRAs usually want 0.5–0.9 to avoid overriding the base composition. Always defer to the LoRA author’s notes when they differ.
Is there a MiniMax H3 LoRA library on this site?
MiniMax3.org publishes a curated LoRA archetype directory on this page, with recommended settings and verified search sources for each category. Per-file detail pages are planned as the community catalog stabilizes.
Try MiniMax H3 with Your LoRA of Choice
Start with a Turbo LoRA for iteration, then swap in a style or character LoRA for the final render. The MiniMax3.org generator lets you preview supported H3 workflows in the browser before you commit GPU time locally.
Try MiniMax H3 Online10 free credits per daily check-in · Up to 30 total
Related MiniMax H3 Guides
Data References
Published 2026-08-15. Last reviewed 2026-08-15. MiniMax3.org is an independent third-party platform and is not affiliated with, endorsed by, or operated by MiniMax. LoRA archetypes on this page describe categories of community- distributed adapters. Search links point to Hugging Face and Civitai; MiniMax3.org does not host third-party LoRA files.