Best List of LoRAs for MiniMax H3 (2026): Turbo, Realism, Camera, Style and Character Swap

MiniMax H3 is the strongest open video model you can run at home right now, and its LoRA scene has exploded since the weights came out. There are LoRAs that cut generation from 25 steps to 4, LoRAs that make skin and faces look real, LoRAs for camera moves, 360 orbits, VHS tape, 1930s cartoons, VR180 and full character swaps. The problem is finding them: they are spread across Hugging Face and Civitai with different file names, triggers and loading rules. I went through the model cards one by one, and this is the list I wish I had when I started. Every LoRA below has a direct link, what it is for, its trigger word, the strength its author recommends, and anything special you need to load it in ComfyUI.

The Full List at a Glance

Here is every LoRA in this guide in one table. Scroll down for the details on each one, including file names and the settings from its model card.

How to Load a MiniMax H3 LoRA in ComfyUI

Most H3 LoRAs load the same way as LoRAs for any other model. If you have not set up H3 yet, start with my MiniMax H3 text-to-video workflow, then come back here.

1. Download the .safetensors file from the model page and put it in ComfyUI/models/loras/. Making a minimax-h3 subfolder keeps things tidy.
2. Add a LoraLoaderModelOnly node between your H3 model loader and the sampler. Video LoRAs for H3 patch the diffusion model only, so you do not need a CLIP input.
3. Set the strength from the table above.
4. Put the trigger word at the start of your positive prompt if the LoRA has one.
5. To stack LoRAs, chain several LoraLoaderModelOnly nodes one after another.

There are three exceptions you should know about. These LoRAs need their own custom nodes, and one of them uses a different folder:

Full or pruned base? Many people run the pruned H3 base because it is smaller. A LoRA trained on the full base can misfire on the pruned one, which is why several repos ship two files. Match them: pruned LoRA with pruned base, full LoRA with full base. If you are on a 16 GB card like my RTX 5080, read my MiniMax H3 out of memory fixes first, because every LoRA adds a little VRAM.

Speed and Acceleration LoRAs

These are the LoRAs that matter most for anyone running H3 locally. A normal H3 render uses around 25 to 28 steps. A distilled acceleration LoRA brings that down to 4 or 8 steps with a small loss in detail, which turns a coffee-break render into something you can iterate on.

1. larryvrh MiniMax H3 Turbo LoRA

Link: huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora

This is the most popular H3 speed LoRA. It runs in 4 to 8 steps, which the author puts at about a 5x speedup, and it keeps H3’s joint video and stereo audio output. It covers text-to-video and image-to-video.

• Recommended file: minimax_h3_turbo_v4_step600_ema.safetensors. The older v1 ckpt850 file is still useful for 4-step clips with heavy motion.
• Strength: 1.0. Raise it to 1.05 to 1.2 if the output looks blurry, or drop to 0.8 to 0.95 if it looks over-sharpened.
• Scheduler: simple.
• Nodes: needs the ComfyUI-MiniMax-H3-Turbo custom nodes (install from Manager or git clone into ComfyUI/custom_nodes/). Put the LoRA in ComfyUI/models/loras/, add the “MiniMax-H3 Turbo LoRA” node between the model loader and the sampler, and use the “MiniMax-H3 Turbo Sampler” with the simple scheduler and 4 or more steps.
• License: Apache 2.0.

2. LightX2V MiniMax H3 Turbo (FL2V and Ref2V)

Link: huggingface.co/lightx2v/Minimax-h3-Turbo (code on GitHub)

LightX2V is the team behind the well-known Wan speed LoRAs, and this repo does the same for H3’s first-last-frame (FL2V) and reference (Ref2V) modes. It is the one to use if you work from images rather than text.

• Files: 4-step and 8-step LoRAs for both modes. The newest are fl2v_turbo_4step v1.3_768p, fl2v 8step v1.1_768p, ref2v_turbo_4step v1.0_768p and ref2v 8step v1.1_768p.
• Format: each LoRA comes as a bf16 file (1.38 GB) and a comfyui_bf16 file (1.96 GB). Take the comfyui one for ComfyUI.
• Loading: no custom node needed. Drop it in ComfyUI/models/loras/ and add it between the loader and the sampler.
• Bonus: the repo also has a 34 GB fp8 full model with the Turbo weights merged in, if you would rather skip the LoRA.

The 768p versions were trained at 768p, so they give cleaner results at that resolution than the original v0.1 files.

3. Alibaba PAI MiniMax H3 Acc LoRAs

Link: huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs

Alibaba’s PAI team (the VideoX-Fun people) released two 8-step acceleration LoRAs: MiniMax-H3-FL2VA-Acc-8Step and MiniMax-H3-Ref2VA-Acc-8Step. They are rank 64, trained with their PDD distillation method, and the card shows them tested at both 4 and 8 NFE. Use a weight of 1.0. The official examples use VideoX-Fun scripts, but the files load as normal LoRAs. License: MiniMax H3 Community License.

4. Taobao PTD LoRAs (4-step and 8-step Ref2VA)

Links: MiniMax-H3-PTD-LoRAs and the RefAsymOPD variant

Released on October 22, these are the newest speed LoRAs on the list. PTD uses a separate adapter for each sampling hop, so a “LoRA” here is really a set: the 8-step pack is 11.1 GB and the 4-step pack is 5.58 GB, both rank 128. They are aimed at Ref2VA. The catch is setup. You need the ComfyUI-PTD and ComfyUI-MiniMaxH3-PTD nodes, the files go in ComfyUI/models/ptd/, ComfyUI must start with --disable-dynamic-vram, and the official route expects the unpruned base model (around 140 GB). The community ComfyUI-PTD-Compact nodes add support for pruned and int8 bases. I would only go here if you already have Turbo working and want the best quality at 4 steps.

5. drbaph Turbo LoRA ComfyUI Conversions

Link: huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI

This repo is a toolbox of speed LoRAs repackaged for ComfyUI. It has Turbo conversions that work on the pruned base, a HyperFlow 8-step LoRA, LightX2V LoRAs resized to a dynamic rank (smaller files, less VRAM), and two experimental ones: TaoMate 3-step and VDN-H3. Use the full files with the full base and the pruned files with the pruned base. License: Apache 2.0. Other community conversions also exist, for example Alberto454, Abiray and Halpak each host a ComfyUI-ready Turbo repack, but I have not checked those cards in detail.

Realism LoRAs

6. fal MiniMax H3 Realism People LoRA

Link: huggingface.co/fal/MiniMax-H3-Realism-People-LoRA

H3 already does people well, but skin can look waxy and lighting a bit too clean. This LoRA from fal pushes it toward natural skin texture and real-camera lighting. It is a single file, h3-realism-people-t2v-i2v-r2v.safetensors (rank 32), that works in text-to-video, image-to-video and reference-to-video.

• Trigger: r34l1sm
• Strength: 1.0, or 0.6 to 0.8 for a lighter touch.
• License: MiniMax H3 Community License.

7. prithivMLmods Facial Realism CloseUp

Link: huggingface.co/prithivMLmods/MiniMax-H3-Facial-Realism-CloseUp

A narrower LoRA for tight face shots from an image plus a text prompt. Use the trigger Facial Realism and pick checkpoint 1800 or 2000, which the author recommends. It is marked experimental, so treat it as something to try on close-up portraits rather than a general fix.

Camera and Motion LoRAs

8. Jojocodex Camera Motion LoRA

Link: huggingface.co/Jojocodex/minimax-h3-Camera-Motion-lora

H3 often ignores camera instructions or picks its own. This LoRA makes it follow them. Start your prompt with camera motion, then describe the move. The card says push-in, pull-back and handheld are the strongest, while pans are weaker. It ships with a library of 55 example prompts, which is worth copying.

• Files: camera_motion_h3_lora_v1_1000_pruned and camera_motion_h3_lora_v1_3000_pruned (for the pruned base).
• Strength: 0.8 to 1.0. Above 1.2 it gets unstable.
• License: Apache 2.0.

For prompt wording that pairs well with it, see my MiniMax H3 prompt guide.

9. Jojocodex Spatial and Physics LoRA

Link: huggingface.co/Jojocodex/minimax-h3-spatial-physics-lora

This one aims at the classic video model failures: hands passing through objects, things floating, bad weight. There is no trigger word; describe the physics you want (“the ball bounces twice and rolls to a stop”). The file is wushu_spatial_physics_clean_3000_pruned. The card suggests 0.3 to 0.5 in general use but 0.8 to 1.0 in its ComfyUI section, so start low and raise it. It can be stacked with Turbo. The author still calls it in testing.

10. Jojocodex Wushu Action v7

Link: huggingface.co/Jojocodex/wushu-action-v7-minimax-h3-fl2va-ref2va-lora

Trained on 924 martial arts clips, this LoRA gives H3 fast, readable fight choreography: kicks, spins and staff work that do not melt into a blur. It works in FL2VA and Ref2VA.

• Trigger: wushu_action
• Strength: 0.6 to 0.8
• Card settings: euler sampler, simple scheduler, 25 steps, CFG 1.0, 832×480, 124 frames.
• License: “other”, so read the card before commercial use.

11. pablodawson 360 Orbit LoRA

Link: huggingface.co/pablodawson/MiniMax-H3-360-Orbit-LoRA

This creates the “bullet time” effect: the scene freezes and the camera circles all the way around it. It uses FL2VA with the same image as both the first and last frame, which is how the loop closes. There is no trigger word. Instead, use the verbatim prompt from the card, which begins “One frozen instant. Only the camera moves. In a continuous 360 orbit.”

• Settings: strength 1.0, 28 steps, 768×768, 73 frames, no guidance, audio off.
• License: MiniMax H3 Community License.

If you have not used first-last-frame mode before, my H3 image-to-video workflow covers it.

Style LoRAs

12. lovis93 STUDIO 1939 Old Animation

Link: huggingface.co/lovis93/studio-1939-old-animation-lora-minimax-h3

Turns H3 into a 1930s animation studio: hand-drawn characters, painted backgrounds and that rubber-hose bounce. Start prompts with gulliv3r, . There are two versions, light (rank 16) and strong (rank 64). Use 1.0 for the full look, or 0.4 to 0.8 to blend it with realism. The card does not state a license, so assume personal use only until the author says otherwise.

13. KennethFal vh5tape VHS LoRA

Link: huggingface.co/KennethFal/vh5tape-vhs-lora-minimax-h3

Gives clips a home-video VHS look with color bleed, tracking noise and soft detail. Because H3 generates audio too, this LoRA also adds tape hiss to the sound, which is a nice touch. The trigger is vh5tape.

• Files: vh5tape.safetensors and vh5tape-comfyui.safetensors (use the comfyui one).
• Strength: 1.0, with a ComfyUI range of 0.7 to 1.2. With Turbo, use 0.4 to 0.6.
• License: MiniMax H3 Community License.

14 to 16. Civitai Style LoRAs: Grimdark 3D, KreaSammelin and KreaIllustrator

Civitai has a growing set of H3 style LoRAs trained on the FL2VA mode. These three have clear cards with triggers and strengths. They are labeled experimental, so expect to tune them.

• Grimdark 3D: dark, gritty 3D render style. Trigger grimdark3dstyle. Strength 1.0 on paper, but the card says it often needs 1.3 to 2.0. Derived from a model by TiwazM.
• KreaSammelin: clean ligne-claire comic look. Trigger kreasammelinstyle. Strength 1.1 (range 1.0 to 1.5). Credits Adel_AI.
• KreaIllustrator: editorial illustration style. Trigger kreaillustratorstyle. Strength 1.0 to 1.2. Credits Adel_AI.

360 and VR Output LoRAs

17. rehan-fal 360 Equirectangular LoRA

Link: huggingface.co/rehan-fal/minimax-h3-360-equirect-lora

Makes H3 output a full 360 degree equirectangular video you can view in a VR headset or upload as a 360 video. Start the prompt with eqr360 followed by the layout sentence from the card. Use strength 0.75. Files are h3-360-equirect-lora-v1.safetensors plus an alternative trained at 896×384. Text-to-video only. License: MiniMax Community License.

18. rehan-fal VR180 Side-by-Side LoRA

Link: huggingface.co/rehan-fal/minimax-h3-vr180-sbs-lora

Generates stereo VR180 video with the left and right eye side by side. Trigger vr180sbs, strength 1.0, and generate at 21:9 and 768P. The card warns against the 8-step “h3-max” route, so run it at normal steps or with a 4-step Turbo LoRA. License: MiniMax Community License.

Character Swap and Character LoRAs

19. Alissonerdx Character Swap v1.1

Link: huggingface.co/Alissonerdx/Minimax-H3-ComfyUI

Swap the person in a video for someone from a reference image while keeping the motion. Version 1.1 (minimax_h3_character_swap_v1.1_ref2va_fro.safetensors, released October 18) is only 80 MB, down from 620 MB for v1.0.

• Settings: strength 1.0, 8 steps with a DMD turbo LoRA, frame count of 17n+5 (for example 107 or 141), and the guide image at frame_idx 0.
• Workflows: two are included. One uses the Shot Planner and needs BFSNodes 1.113 or newer; the other uses native nodes only.

20. akatz-ai Character Swap LoRA V1.5

Link: huggingface.co/akatz-ai/MiniMax-H3-Character-Swap-LoRA

The other big character swap LoRA, also video-to-video on Ref2VA. Use h3_character_swap_v1.5.safetensors. In the prompt you refer to the source clip as <Video 1> and the new character as <Picture 1>.

• Settings: strength 1.0, 8 steps, CFG 1, er_sde sampler with the normal scheduler, together with an 8-step Turbo LoRA.
• Workflow: H3-Character-Swap-Relay-V2.json, which needs H3 Relay 2.0.
• Try it online: there is a Hugging Face Space, linoyts/minimax-h3-character-swap.
• License: MiniMax H3 Community License.

Which swap LoRA is better? Alissonerdx is far smaller and has a native-only workflow, which helps on 16 GB cards. akatz-ai has the easier prompt format and an online demo. Try both on the same clip. My H3 reference-to-video workflow is a good base for either.

21. masafy Character LoRA

Link: huggingface.co/masafy/minimax-h3-masafy-lora

A single-character LoRA on Ref2VA. It is most useful as a working example of what a character LoRA for H3 looks like. Trigger masafy_character plus a description of the appearance, checkpoint 500 at strength 0.8, and the card reports about 9.3 GB of VRAM.

Streaming, Prompt Rewriter and Collections

22. mvp-lab RAVEN Streaming LoRA

Link: huggingface.co/mvp-lab/MiniMax-H3-RAVEN-Streaming-LoRA

RAVEN turns H3 into a causal, streaming text-to-video model that generates the clip chunk by chunk at 4 NFE. It is rank 128 and produces up to 192 frames at 1376×768. The ComfyUI-MiniMax-H3-RAVEN-Streaming nodes are built to fit a 24 GB budget. They expect the LoRA at ComfyUI/models/loras/minimax_h3_raven_streaming_lora_4nfe_preview.safetensors and the full, non-pruned minimax_h3_fl2va_bf16.safetensors in ComfyUI/models/diffusion_models/. Nothing downloads automatically. It is a preview release, so expect rough edges, but it is the most interesting route to longer H3 clips.

Bonus: LightX2V Prompt Rewriter LoRA

Link: huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA

Not a video LoRA. It sits on top of Qwen3.6-27B and rewrites your short prompt into the long, structured style H3’s text-to-video-with-audio mode likes. Handy if your prompts come out flat.

Bonus: 110-LoRA Mirror Collection

Link: huggingface.co/Alex995647/loras-minimax-h3

This repo mirrors 110 H3 LoRAs (63.5 GB) from Civitai and Hugging Face in one place. It is useful for browsing, but licenses vary by LoRA, so check the original page before using any of them in public work.

Not LoRAs: you will see Singularity V2 and FastH3 Omni mentioned alongside these. They are full fine-tuned checkpoints, not LoRAs, so they replace your base model instead of stacking on it.

Stacking LoRAs: What Works

The most useful combination is one speed LoRA plus one creative LoRA. Here is how to set it up without breaking the output.

• Speed LoRA first, at 1.0. Put the Turbo LoRA closest to the model loader and keep it at its recommended strength.
• Creative LoRA lower. Distilled LoRAs amplify other LoRAs. The VHS card explicitly drops from 1.0 to 0.4 to 0.6 with Turbo, and the same rule of thumb works for most style LoRAs.
• Match the mode. An FL2VA LoRA stacked on a T2V workflow will do little. Check the Modes column above.
• Do not stack two speed LoRAs. Pick Turbo or LightX2V or PAI, not two at once.
• Two creative LoRAs at most. Camera Motion plus a style LoRA works well; three or more tends to fight.
• Change one thing at a time. Fix the seed, then adjust strength in 0.1 steps.

Once you have a clip you like, you can smooth it to 60 fps with my MiniMax H3 RIFE workflow, or fix a single bad section with the H3 video section editor instead of rendering everything again.

Licenses: Can You Use Them Commercially?

There are three groups. Apache 2.0 LoRAs (larryvrh Turbo, drbaph conversions, Camera Motion, Spatial and Physics) are the most permissive for the LoRA itself. Many others use the MiniMax H3 Community License, the same license as the base model, so the base model’s terms apply. A few say “other” or nothing at all (Wushu Action, STUDIO 1939), and for those you should ask the author before commercial use. Remember that whatever the LoRA says, your output is still covered by the H3 base model license.

Train Your Own MiniMax H3 LoRA

If nothing on this list fits, training your own is realistic. fal has a clear guide to training a MiniMax H3 LoRA, and ostris ai-toolkit has a minimax_h3 extension for local training. Style and camera LoRAs need the fewest clips; the Wushu LoRA used 924, but most style LoRAs use far fewer. Keep each clip short, caption the motion as well as the look, and pick a rare trigger word like the ones above.

FAQ

What is the best LoRA for MiniMax H3?

For most people, a speed LoRA: larryvrh Turbo for text and image to video, or LightX2V Turbo for first-last-frame and reference modes. After that, fal Realism People is the best all-round quality upgrade.

Where do MiniMax H3 LoRAs go in ComfyUI?

In ComfyUI/models/loras/, loaded with a LoraLoaderModelOnly node. The exceptions are the Taobao PTD adapters, which go in ComfyUI/models/ptd/, and LoRAs that need their own custom nodes such as larryvrh Turbo and RAVEN.

Do LoRAs work with the pruned H3 model?

Use LoRA files made for the pruned base. Repos such as drbaph and Jojocodex label their pruned files. A full-base LoRA on the pruned model can give weak or broken results.

Can I use LoRAs on a 16 GB GPU?

Yes. Each LoRA adds some VRAM, so use the smaller files (Alissonerdx v1.1 is 80 MB, drbaph’s dynamic-rank resizes are smaller than the originals) and apply the memory fixes from my 16 GB guide. Avoid PTD and RAVEN on 16 GB for now.

Can I use Wan or LTX LoRAs on MiniMax H3?

No. LoRAs are tied to the architecture they were trained on. Use only LoRAs made for MiniMax H3.

Do these LoRAs affect the audio?

Some do. Turbo keeps the joint audio, and vh5tape deliberately adds tape hiss. The 360 Orbit card turns audio off. If audio breaks after adding a LoRA, try lowering its strength.

Final Thoughts

If you only install two LoRAs, make them a Turbo LoRA and fal Realism People. That combination turns MiniMax H3 from a slow, impressive demo into something you can actually iterate on at home. From there, Camera Motion and 360 Orbit give you shots that are hard to prompt, the character swap LoRAs open up real editing work, and the style LoRAs are pure fun. New H3 LoRAs are appearing every week, and I will keep this list updated as I verify them.

More MiniMax H3 guides:

• Best MiniMax H3 Text-to-Video Workflow for ComfyUI
• Best MiniMax H3 Image-to-Video Workflow for ComfyUI
• Best MiniMax H3 Reference-to-Video Workflow for ComfyUI
• MiniMax H3 ComfyUI Reference to Image
• MiniMax H3 Prompt Guide
• MiniMax H3 Out of Memory in ComfyUI: 16 GB GPU Fixes
• Best MiniMax H3 RIFE Workflow in ComfyUI
• Best MiniMax H3 Video Section Editor in ComfyUI
• Best Local AI Video Models in 2026

Sources

• Hugging Face: larryvrh/MiniMax-H3-Turbo-Lora
• Hugging Face: lightx2v/Minimax-h3-Turbo
• Hugging Face: alibaba-pai/MiniMax-H3-Acc-LoRAs
• Hugging Face: taobao-content-ai/MiniMax-H3-PTD-LoRAs
• Hugging Face: drbaph/MiniMax-H3-Turbo-Lora-ComfyUI
• Hugging Face: fal/MiniMax-H3-Realism-People-LoRA
• Hugging Face: Jojocodex Camera Motion LoRA
• Hugging Face: pablodawson/MiniMax-H3-360-Orbit-LoRA
• Hugging Face: akatz-ai/MiniMax-H3-Character-Swap-LoRA
• Hugging Face: Alissonerdx/Minimax-H3-ComfyUI
• fal: How to train a LoRA for MiniMax H3