Uncensored Local AI Video Generation in ComfyUI (2026): Wan 2.2, LTX-2.5, MiniMax H3 and Kandinsky 6

Cloud video generators are even stricter than image generators. A fight scene for a short film, a monster reveal for a horror trailer, a war re-enactment or a stunt sequence can all be blocked or quietly toned down. Open video models running in ComfyUI on your own GPU have no such filter, which is why “uncensored” local video is one of the most searched AI topics of 2026. But as with images, the newest video models read your prompt with a full chat LLM, and that LLM can still weaken a prompt it does not like. This guide covers the best open video models for creative freedom as of November 2026 (Wan 2.2, LTX-2.5, MiniMax H3 and Kandinsky 6.0), which text encoder each one uses, the Heretic and abliterated encoders you can swap in, and the exact files, nodes and ComfyUI folders you need.

Why Local Video Models Still Hold Back

ComfyUI never filters prompts or output. When a local video model ignores part of your prompt, there are three possible reasons:

1. The training data. The model simply has not seen enough of something (a specific fighting style, creature, camera move) to animate it. Only a LoRA or fine-tune fixes this.
2. The text encoder. LTX-2.x and MiniMax H3 use instruction-tuned LLMs (Gemma and Qwen3-VL) as encoders, and those carry refusal behaviour. The DeepNeuralNerd card for LTX-2.5 puts it well: standard Gemma “has residual refusals that can still gut perfectly innocent prompts”. A Heretic or abliterated encoder removes that.
3. Prompt enhancers. Workflows that expand your prompt with an LLM (including image-to-video “prompt enhance” nodes) inherit that LLM’s refusals. Use a local, uncensored LLM for this step (see our uncensored local LLM guide).

We explained the encoder side in detail for images in Uncensored Local AI Image Generation in ComfyUI. The same idea applies to video, with bigger files.

The Best Open Video Models Compared

Wan 2.6, Wan 3.0 and FLUX 3 Video are API-only, so they are not covered here.

Wan 2.2: The Most Open Option

Wan 2.2 from Alibaba remains the easiest model to use without restrictions, for a simple reason: its encoder is umT5-XXL, a T5 model that has no chat training and therefore no refusal behaviour at all. It also has the largest community LoRA library of any open video model and an Apache 2.0 licence that allows commercial use. Its weakness is age: it has no audio and its clips are shorter and lower resolution than H3 or LTX-2.5.

Files (from Comfy-Org/Wan_2.2_ComfyUI_Repackaged):

• Text-to-video, two experts: ComfyUI/models/diffusion_models/wan2.2_t2v_high_noise_14B_fp8_scaled.safetensors and ComfyUI/models/diffusion_models/wan2.2_t2v_low_noise_14B_fp8_scaled.safetensors
• Image-to-video, two experts: ComfyUI/models/diffusion_models/wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors and ComfyUI/models/diffusion_models/wan2.2_i2v_low_noise_14B_fp8_scaled.safetensors
• Small model for 8 to 12 GB: ComfyUI/models/diffusion_models/wan2.2_ti2v_5B_fp16.safetensors
• Text encoder: ComfyUI/models/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors
• VAE: ComfyUI/models/vae/wan_2.1_vae.safetensors for the 14B models, ComfyUI/models/vae/wan2.2_vae.safetensors for the 5B model

Load the official templates from Workflow > Browse Templates > Video. On 12 to 16 GB cards, the GGUF versions of the 14B experts (via ComfyUI-GGUF) and the LightX2V 4-step distill LoRAs in ComfyUI/models/loras/ are the usual way to get usable speed.

LTX-2.5 With a Heretic Gemma Encoder

LTX-2.5 from Lightricks (released August 11, 2026, 22B) generates video with synced audio and has the most complete set of official tools: upscalers, IC-LoRAs and distilled checkpoints. Its encoder is Gemma 4 12B with LTX’s own projection layers bundled into the same file, which is why you cannot just drop in any uncensored Gemma. The replacement has to keep those projections.

Heretic encoder: DeepNeuralNerd/Gemma-4-12B-it-uncensored-heretic-DeepNeuralNerd-LTX_2.5_ComfyUI. It is built from llmfan46’s Gemma 4 12B Heretic and keeps the original LTX-2.5 video and audio projection weights. It comes in bf16 and INT8 ConvRot. The author is clear that it improves prompt fidelity but “does not uncensor the video model itself”, and it is a community conversion with no formal benchmark.

Files and folders (official files are in the gated Lightricks/LTX-2.5 repo; accept the licence first):

• Model: ComfyUI/models/checkpoints/ (INT8 ConvRot distilled 21.5 GB, NVFP4 distilled 18.7 GB, or a community GGUF Q4_K_M 15.7 GB for 16 GB cards)
• Heretic encoder: ComfyUI/models/text_encoders/Gemma-4-12B-it-uncensored-heretic-DeepNeuralNerd-LTX_2.5_ComfyUI-bf16.safetensors (or the INT8 version), selected in place of the official Gemma 4 12B encoder (official INT8 is 15.4 GB)
• VAEs: ComfyUI/models/vae/ (video and audio VAE)
• Latent upscalers: ComfyUI/models/latent_upscale_models/
• LoRAs: ComfyUI/models/loras/ (many LTX-2.3 style LoRAs load on 2.5, but use the 2.5 versions of the official IC-LoRAs; full setup in our LTX-2.5 16 GB guide)

Nodes: install ComfyUI-LTXVideo (Manager name: ComfyUI-LTXVideo) for the example workflows, and ComfyUI-GGUF if you use GGUF files. Manual install:

cd ComfyUI/custom_nodes
git clone https://github.com/Lightricks/ComfyUI-LTXVideo
cd ComfyUI-LTXVideo
pip install -r requirements.txt

LTX-2 and LTX-2.3 (Gemma 3 12B)

If you are still on LTX-2 or 2.3, use DreamFast/gemma-3-12b-it-heretic-v2. It was made with Heretic 1.3.0 over 200 trials: refusals fell from 100/100 to 8/100 with a KL divergence of 0.08. The ComfyUI files keep the vision weights needed for image-to-video prompt enhancement.

• Files: FP8 or INT8 13 GB, NVFP4 7.8 GB, INT4 7.7 GB, bf16 23 GB. Put it in ComfyUI/models/text_encoders/ and select it in the LTXAVTextEncoderLoader node.
• GGUF (Q4_K_M 6.8 GB): load with DualCLIPLoader (GGUF), with CLIP 1 set to the Gemma GGUF and CLIP 2 set to the dev embedding connectors from Kijai’s LTXV2_comfy repo.
• INT8 needs ComfyUI 0.27+, INT4 needs 0.30+.

MiniMax H3 With an Abliterated Encoder

MiniMax H3 (33B) gives the best motion and audio of any open model, and we have a full series on it, starting with our 16 GB setup guide and the H3 Extender workflow. Its encoder is Qwen3-VL 32B trimmed to the 50 layers H3 uses: officially qwen3vl_32b_minimax_h3_int8_convrot.safetensors (27.1 GB) or qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (15.7 GB) for RTX 50-series.

Abliterated encoder: pottokao/MiniMax-H3-TextEncoder-16GB-abliterated-GGUF (October 21). It takes the abliteration from llmfan46’s Qwen3-VL 32B Heretic, where the only changed tensors are the attention outputs in layers 31 to 39, and keeps the official H3 tensor layout and vision tower.

• Files: ...abliterated-11.5G_vis.gguf (12.65 GB), ...-12.5G_vis.gguf (13.70 GB), ...-13.6G_vis.gguf (14.80 GB)
• For a 16 GB card use the 11.5G file. The author ran full clips on an RTX 5070 Ti at about 75 seconds per clip with a GPU peak near 15 GB. The two larger files ran out of memory on 16 GB at default settings.
• Folder: ComfyUI/models/text_encoders/
• Loader: CLIPLoaderGGUF from ComfyUI-GGUF with type minimax
• The author measured about 8% drift from the official encoder’s output and makes no claim about whether that helps a given video. Test with the same seed.

Everything else stays the same as the official H3 workflow: the model in ComfyUI/models/diffusion_models/, video and audio VAEs in ComfyUI/models/vae/, and Turbo or LoRA files in ComfyUI/models/loras/. Note the H3 licence excludes some territories (including the US, EU, UK and South Korea) and needs authorisation above $20M yearly revenue.

Kandinsky 6.0: New Open Model With Audio

Kandinsky Lab released Kandinsky 6.0 on October 6, 2026 under the MIT licence: a 29B Pro and a 3B Lite model, each with a distilled 10-step checkpoint, generating 5-second 24 fps clips with synchronized 44 kHz audio and lip-sync, plus a super-resolution model for Full HD. It encodes prompts with Qwen2.5-VL plus a CLIP model. No dedicated uncensored encoder for it has been published yet, but the MIT licence and the small 3B Lite make it a strong option for experimenting on mid-range cards.

Setup: in ComfyUI Manager install kandinsky6 and kandinsky6-sr, restart, and open Workflow > Browse Templates > kandinsky6. The extension’s Download models button puts every file in the right folder.

LoRAs for Action, Horror and Cinematic Shots

Because no encoder can teach the model new subjects, LoRAs are where real creative freedom comes from. The most useful categories for film-makers:

• Camera and motion LoRAs: orbit, crash zoom, handheld, drone and bullet-time moves (common for Wan 2.2 and H3; see our best MiniMax H3 LoRAs).
• Action and VFX LoRAs: explosions, fire, sword fights, sci-fi weapons, slow motion (LTX-2.5 has an official Slow Motion Control I2V LoRA).
• Horror and dark-fantasy style LoRAs: film grain, VHS, gothic lighting, creature transformations.
• Character consistency LoRAs trained on your own original characters.

Put them in ComfyUI/models/loras/ and add a LoraLoaderModelOnly node after the model loader (for Wan 2.2 A14B, add one to each expert). Always check that a LoRA matches your exact model version.

VRAM Guide

For more memory tricks, see ComfyUI launch flags (for example --reserve-vram 1.5) and gpu-unlock for RTX 30/40.

Prompting Tips for Video

• Describe motion, not just the scene. “The knight swings the sword in a wide arc, sparks fly, the camera pulls back” gives the model something to animate.
• Name the camera. Shot size, lens and movement (“low-angle tracking shot, 35 mm”) matter more than adjectives.
• Same seed, two encoders. Compare the stock and Heretic encoders on one seed. If the clips match, the limit is the model, not the encoder.
• Keep enhancers local. If a workflow rewrites prompts with an LLM, point it at a local model.

Responsible Use and the Law

• Never make sexual content involving minors, or videos of real people in sexual, violent or humiliating situations without consent. Deepfake video is now a specific crime in many countries, including the US, UK, EU and India.
• Do not create fake news or impersonation videos. Realistic video of real public figures saying or doing things they did not can be fraud or defamation.
• Label AI video when you publish it, and follow each platform’s synthetic-media rules.
• Respect licences. Wan 2.2 and Kandinsky 6.0 are permissive; LTX-2.5 and MiniMax H3 have revenue and territory conditions.
• Keep ComfyUI private. Do not expose it publicly without authentication.

Troubleshooting

• LTX-2.5 errors about missing projection weights: you loaded a plain Gemma 4 12B Heretic. Use the DeepNeuralNerd LTX_2.5 file, which includes the projections.
• H3 encoder out of memory on 16 GB: use the 11.5G file, or launch with --reserve-vram 1.5.
• “unknown architecture” with a GGUF encoder: update ComfyUI-GGUF and the gguf Python package.
• No visible difference: normal for most prompts. The encoder only matters where the stock one was weakening a prompt.
• More fixes: ComfyUI errors and fixes.

FAQ

What is the best uncensored local video model in 2026?

Wan 2.2 is the most open, because its T5 encoder never refuses. For quality with audio, MiniMax H3 or LTX-2.5 with a Heretic or abliterated encoder.

Does Wan 2.2 have a content filter?

No. Running locally in ComfyUI there is no filter, and its umT5 text encoder has no refusal training.

What text encoder does LTX-2.5 use?

Gemma 4 12B with LTX’s projection layers bundled. Older LTX-2 and 2.3 use Gemma 3 12B.

Can I use any uncensored Gemma with LTX-2.5?

No. The file must contain LTX-2.5’s video and audio projection weights, like the DeepNeuralNerd conversion does.

Will an uncensored encoder make a video model generate anything?

No. It only stops the encoder from diluting your prompt. The video model can still only produce what it learned; new subjects need LoRAs.

Which local video model is best for 16 GB VRAM?

MiniMax H3 with the 11.5G abliterated encoder and block swap, or Wan 2.2 14B as GGUF with the 4-step LoRAs.

Final Thoughts

Local video is where open models give you the most freedom over cloud services. Wan 2.2 is already unrestricted thanks to its T5 encoder, while LTX-2.5 and MiniMax H3 become far more faithful to your prompt with a Heretic or abliterated encoder. Pick by VRAM, add LoRAs for the motion and styles you need, test with the same seed, and keep your projects legal and labelled.

Related guides:

• Uncensored Local AI Image Generation in ComfyUI
• Best Uncensored Local LLMs in 2026
• Best Local AI Video Models in 2026
• Best LoRAs for MiniMax H3
• ComfyUI Model Folder Cheat Sheet

Sources

• DeepNeuralNerd LTX-2.5 Heretic encoder and DreamFast Gemma 3 12B Heretic v2 model cards
• pottokao MiniMax H3 abliterated encoder
• Lightricks/LTX-2.5 and LTX 2.5 ComfyUI setup guide
• ComfyUI Wiki: Kandinsky 6.0
• Fuser: best open-source video models and licences (updated Oct 23, 2026)
• Comfy-Org Wan 2.2 repackaged files
• Heretic Docker: abliterating text encoders for video models
• Hugging Face model search for LTX, Gemma 3 12B and MiniMax H3 encoder variants (checked November 2, 2026)