LTX 2.5 & LTX 2.3 Model Downloads

Community Q3_K_S GGUF quantization of the LTX 2.5 distilled transformer. Not gated — no HuggingFace license click-through required. Smallest quant in this line; expect more quality loss than Q4_K_M/Q6_K. Requires the ComfyUI-GGUF custom node to load. Place in models/checkpoints/.

Community Q4_K_M GGUF quantization of the LTX 2.5 distilled transformer — the commonly recommended balance point between size and quality in the GGUF quant family. Not gated. Requires the ComfyUI-GGUF custom node to load. Place in models/checkpoints/.

Community Q5_K_M GGUF quantization of the Gemma 4 text encoder — at 9.51GB, roughly the same footprint as LTX 2.3's Gemma 3 FP4 encoder (9.5GB), which is what makes a 16GB LTX 2.5 setup possible at all. The highest-quality of the three GGUF encoder quants; Q4_K_M (8.41GB) and Q2_K (5.96GB) trade fidelity for more transformer headroom. The repo inherits LTX 2.5's gating — sign in and accept the license. Requires the ComfyUI-GGUF custom node to load. Place in models/text_encoders/.

Community Q4_K_M GGUF quantization of the Gemma 4 text encoder — 8.41GB, 1.1GB smaller than Q5_K_M. That headroom is what lets a 16GB card pair the encoder with a Q4_K_M or Q5_K_M transformer instead of being forced down to Q3_K_S. Projection layer is bundled, so no separate text-projection file. The repo inherits LTX 2.5's gating — sign in and accept the license. Requires the ComfyUI-GGUF custom node. Place in models/text_encoders/.

Community Q2_K GGUF quantization of the Gemma 4 text encoder — 5.96GB, the smallest Gemma 4 encoder available for LTX 2.5. Q2_K is an aggressive 2-bit quant; expect weaker prompt adherence than Q4_K_M/Q5_K_M. Use it only when the encoder must be as small as possible and Q4_K_M still doesn't fit your budget. Projection layer bundled. The repo inherits LTX 2.5's gating. Requires the ComfyUI-GGUF custom node. Place in models/text_encoders/.

Official INT8 convrot quantized dev transformer — the first-party ComfyUI quant for LTX 2.5 (filename literally says 'comfy'). Unlike LTX 2.3, there is no separate FP8 build and no Kijai fork for 2.5; this is the smallest official dev checkpoint. Pairs with a ~15.37GB+ Gemma 4 text encoder, so a fully-official pipeline needs well above 24GB. Repo is gated.

Official INT8 convrot quantized distilled transformer for LTX 2.5 — the first-party ComfyUI quant. Fastest official 2.5 inference path. Pairs with a ~15.37GB+ Gemma 4 text encoder, so a fully-official pipeline needs well above 24GB. Repo is gated.

Official NVFP4 quantized distilled transformer — the smallest official 2.5 checkpoint. Native NVFP4 matmul on Blackwell (RTX 50xx); falls back to a slower path on older GPUs. Repo is gated.

Official INT8 quantized Gemma 4 12B text encoder for LTX 2.5, projection bundled in. This is the smallest official text encoder — at 15.37GB it's already larger than LTX 2.3's smallest encoder (9.5GB), which is why 16GB cards can't run an all-official LTX 2.5 pipeline. Place in models/text_encoders/. Repo is gated.

Community Q6_K GGUF quantization of the LTX 2.5 distilled transformer — closer to full quality than Q4_K_M at a larger size. Not gated. Requires the ComfyUI-GGUF custom node to load. Place in models/checkpoints/.

Community Q8_0 GGUF quantization of the LTX 2.5 distilled transformer — the highest-fidelity GGUF quant, close to the official int8-convrot checkpoint in size. Not gated. Requires the ComfyUI-GGUF custom node to load. Place in models/checkpoints/.

Official full-precision BF16 dev transformer for LTX 2.5. Source weights every quantized 2.5 variant derives from. Requires 48GB VRAM or sequential offloading on 32GB. Repo is gated — sign in to HuggingFace and accept the license before downloading.

Official full-precision BF16 distilled transformer for LTX 2.5. Requires 48GB VRAM or sequential offloading on 32GB. Repo is gated — sign in to HuggingFace and accept the license before downloading.

Full BF16 Gemma 4 12B text encoder for LTX 2.5, with the projection layer bundled into this single file — LTX 2.3 needed a separate text-projection file, 2.5 does not. Required for every LTX 2.5 workflow. Place in models/text_encoders/. Repo is gated.

Video VAE that decodes LTX 2.5 latents into frames. Required for all LTX 2.5 ComfyUI workflows. Place in models/vae/. Repo is gated.

Convolutional video VAE variant shipped alongside the standard LTX 2.5 video VAE — no equivalent existed in LTX 2.3. Place in models/vae/. Repo is gated.

Audio VAE for LTX 2.5 audio-video generation. Place in models/vae/. Repo is gated.

Official distillation LoRA for LTX 2.5, applied on top of the dev model to get distilled-style fast inference. Rank 450 — up from rank 384 on LTX 2.3's distillation LoRA. Place in models/loras/. Repo is gated.

Latent spatial upscaler for two-stage LTX 2.5 pipelines. Doubles spatial resolution in latent space before VAE decode. Place in models/latent_upscale_models/. Repo is gated.

Latent temporal upscaler for LTX 2.5 — doubles frame count for smoother motion in a two-stage pipeline. Place in models/latent_upscale_models/. Repo is gated.

Small model patch shipped only with LTX 2.5 — no equivalent exists for LTX 2.3. Place in models/model_patches/. Repo is gated.

Strength 1.0 — the 2.5 weights ship pre-scaled, unlike the 2.3 version's recommended 1.4. Use the two-part 'Reference sheet: … / Generated video: …' prompt. A generic LoRA loader that ignores the reference path will not apply the conditioning; use an IC-LoRA workflow.

Reference-sheet control IC-LoRA for LTX 2.5: conditions generation on a single composite image inventorying a scene's characters, props, and location so they stay consistent in the output. The most-downloaded of the 2.5 Creative Lab adapters. Trained at one bucket only (768×448, 121 frames, 24 fps). Gated repo. Place in models/loras/.

No trigger word — describe the desired empty scene, not a 'remove X' instruction. Strength 1.0. Dimensions divisible by 32, frame counts 8k+1. Not for subjects that fill most of the frame.

VFX IC-LoRA for LTX 2.5 that removes people and other dynamic subjects from a source video and reconstructs the empty background. Trained at 1024×576 / 576×1024, 49 frames @ 25 fps; validated at 1920×1088. The only 2.5 Creative Lab adapter shipped as a 1.0 rather than 0.9. Gated repo. Place in models/loras/.

DEBLUR trigger in the two-part caption; strength 1.0, lower toward 0.8 if it halos. Shipped checkpoint is training step 1000 of a planned 1500 — the card says the ~800–1000 range is the sweet spot. Motion blur, compression repair, and denoising are out of scope.

Restoration IC-LoRA for LTX 2.5 that recovers sharpness from out-of-focus footage by conditioning on the blurry clip, preserving subject, framing, and geometry. Whole-frame, no mask. Trained at 960×544, 121 frames @ 24 fps. Gated repo. Place in models/loras/.

ENHANCE QUALITY trigger, strength 1.0, (frames−1) % 8 == 0. Compression artifacts only — blur, grain, upscaling, and colorization are explicitly out of scope.

Restoration IC-LoRA for LTX 2.5 that removes compression artifacts — macroblocking, chroma bleed, ringing, banding — from low-bitrate video while keeping identity, framing, and geometry. Trained and validated at 960×544×121 @ 24 fps, both orientations. Gated repo. Place in models/loras/.

COLORIZE trigger; describe the same scene in both caption halves, changing only the color language. Strength 1.0. Card's production recipe is the distilled pipeline, stage-1-only on a 2× canvas — no CFG, no negative prompt (fixed sigmas).

Restoration IC-LoRA for LTX 2.5 that adds natural color to grayscale or desaturated video while leaving identity, framing, and geometry untouched. Trained and validated at 960×544×121 @ 24 fps, both orientations. Gated repo. Place in models/loras/.

ADD WATER trigger in every prompt. Trained on real water; other liquids generalize loosely. Strengths ≥ 1.5 can warp faces and fine detail.

VFX IC-LoRA for LTX 2.5 that adds water — rivers, surf, rain, waterfalls, floods, splashes, spray, wet-surface specularities — while keeping the subject's identity, clothing, pose, framing, and background geometry identical to the reference. Gated repo. Place in models/loras/.

Strength 1.0, 30 steps, guidance 3.0–4.0 (lower = brighter, higher = darker night). Exterior daylight footage; indoor and artificial-light scenes are out of scope, and clips much longer than ~4 s can drift.

Relighting IC-LoRA for LTX 2.5 that re-renders a daytime video as the same shot at night, preserving composition, framing, camera movement, and scene structure. Trained at 768×448 / 448×768, 97 frames @ 24 fps. Gated repo. Place in models/loras/.

Pixel-space spatial upscaler IC-LoRA for LTX 2.5, published as its own separate gated HuggingFace repo. Attaches to the dev/distilled transformer. Place in models/loras/.

CINEMAGRAPH_MOTION trigger in the positive prompt. Card recommends strength 0.5–1.5 with a 1.0–1.2 sweet spot (narrower than the 2.3 version's 0.7–3.0), STG stg_v scale 1.0. Static camera, single moving element only.

Style LoRA for LTX 2.5 (not an IC-LoRA — no reference video) that turns a still image into a smooth looping cinemagraph with one moving element and a frozen, tripod-locked frame. Gated repo. Place in models/loras/.

Needs a pipeline that exposes the speed value: keep frame_rate at 24 and set speed = 1.0 (real-time), 0.5 (2×), 0.2 (5×), 0.1 (10×), 0.05 (20×). No trigger word — write a normal I2V caption and do not add 'slow motion' or fps wording. Through a generic LoRA loader that ignores speed you get a plain LoRA and none of the effect.

Image-to-video LoRA that turns physical motion speed into a knob — real-time through 40× slow-motion — while the output stays at 24 fps playback. Standard LoRA, no reference video, no preprocessing. The one LTX 2.5 Creative Lab repo that is not gated, and it ships an official ComfyUI workflow (LTX-2.5_I2V_Speed_Control_flow.json). Place in models/loras/.

Official FP8 from Lightricks. Alternative to Kijai's FP8 dev model.

Official FP8 dev model from Lightricks. 29.1GB, runs on 16GB VRAM.

Official FP8 distilled from Lightricks. Alternative to Kijai's FP8.

Official FP8 distilled model from Lightricks. 8 steps, CFG=1.

Best choice for 16GB VRAM. Latest v1.1 FP8 distilled. Requires RTX 40xx+ for fp8 matmuls.

FP8 quantized v1.1 distilled by Kijai. Best for 16GB VRAM. 8 steps, CFG=1.

RTX 30xx workaround — use this when your GPU lacks RTX 40xx-style FP8 matmul support. Same VRAM as fp8_scaled.

MXFP8 block-32 quantized distilled 1.1 by Kijai. Use on RTX 30xx GPUs that cannot run standard FP8 scaled matmul.

Use this (not distilled) if you want to apply LoRA weights on 16GB VRAM.

FP8 quantized dev model by Kijai. Runs on 16GB VRAM. Supports LoRA. Place in models/checkpoints/.

Alternative FP8 quantization for dev. Use when fp8_input_scaled has compatibility issues.

FP8 scaled dev model by Kijai. 16GB VRAM. Supports LoRA. Place in models/checkpoints/.

RTX 30xx workaround for dev/quality path — use when standard FP8 matmul is unsupported. Supports LoRA.

MXFP8 block-32 dev model by Kijai. 16GB VRAM. Supports LoRA. Use on RTX 30xx GPUs that cannot run standard FP8 scaled. Place in models/checkpoints/.

Best size/quality tradeoff on RTX 50xx. On older GPUs it falls back to slow paths — prefer FP8 there.

NVFP4 quantized dev model. Native nvfp4 matmul on Blackwell GPUs (RTX 50xx). 21.7GB.

Best low-VRAM distilled option for RTX 30xx. Smaller than MXFP8 (~21.5 GB) and runs on INT8 tensor cores every Ampere+ card has.

INT8 convrot quantized distilled 1.1 by Kijai. Runs on RTX 30xx (Ampere INT8 tensor cores) — the smallest distilled transformer that keeps near-FP8 quality without needing FP8 matmul.

Low-VRAM dev/quality path for RTX 20xx/30xx — INT8 tensor cores give real speedup where FP8 matmul is unavailable and MXFP8 only falls back to BF16.

INT8 convrot quantized dev model by Kijai. 16GB VRAM, supports LoRA, runs on RTX 20xx/30xx INT8 tensor cores. Place in models/checkpoints/.

Pair with the dev model (dev FP8 on 16 GB, BF16 dev on 32 GB+) and set the sampler to 8 steps, CFG = 1. At rank 384 / 7.6 GB it is the heaviest, highest-fidelity distillation LoRA — prefer it when you want maximum distilled quality and have the VRAM headroom.

Official Lightricks rank-384 distillation LoRA, v1.1. Applied on the dev model, it converts dev-style inference (many steps, CFG > 1) into distilled-style (8 steps, CFG = 1) — the official high-rank counterpart to Kijai's dynamic-rank LoRA. Place in models/loras/.

Pair with the dev FP8 model. Load as LoRA in ComfyUI models/loras/.

Distilled LoRA v1.1 by Kijai. Use with the dev model for distilled-quality output on 16GB VRAM.

Pair with the dev FP8 model when running image-to-video or other input-conditioned workflows. Use the standard Kijai dynamic-rank LoRA for T2V instead.

Experimental community distilled LoRA v1.1 by TenStrip. The 'cond-safe' variant zeroes cross-attention bridges, adaln/scale-shift tables, gate logits, and prompt scale-shift — making it better suited for I2V and input-conditioned workflows than the standard dynamic LoRA. Place in models/loras/.

v1.0 dynamic LoRA. Use v1.1 dynamic LoRA (rank 111) for latest quality.

Dynamic rank LoRA (avg rank 105) by Kijai. Pair with dev model. Place in models/loras/.

Add on top of your LTX 2.3 base model to improve audio-video sync and motion stability, especially for audio-conditioned and talking-head workflows. 617 MB BF16. As a behavioral (RL) LoRA it does not change your sampler steps/CFG the way a distillation LoRA does.

Reinforcement-learning quality LoRA for LTX 2.3 from the OmniNFT research line. Targets audio-video synchronization, lip-sync, motion coherence, and temporal stability rather than visual style — reported to cut audio-video DeSync from 0.569 to 0.269 on JavisBench. Place in models/loras/.

Enable sequential offloading in ComfyUI settings. Uses latest v1.1 official weights.

Official v1.1 distilled model runnable on 24GB with sequential offloading enabled in ComfyUI.

For 32GB VRAM: enable Sequential Offloading in ComfyUI settings (file is 46GB). For fastest inference, use the FP8 distilled variant instead.

Official v1.1 distilled model. 8 steps, CFG=1. Latest release from Lightricks. Requires sequential offloading on 32GB — file is 46GB.

Use when FP8 matmul (RTX 40xx+) is unavailable. 44GB — requires 48GB VRAM or sequential offloading on 32GB.

BF16 distilled 1.1 transformer-only by Kijai. For RTX 30xx or GPUs without FP8 support. Requires 48GB+ or sequential offloading on 32GB.

Best for LoRA training and fine-tuning. 42GB — enable Sequential Offloading on 32GB cards.

Full BF16 dev model. Flexible and trainable. 42GB — requires 48GB VRAM or sequential offloading on 32GB.

Use when training LoRA with full BF16 precision. 44GB — enable Sequential Offloading on 32GB cards.

BF16 dev transformer-only by Kijai. 44GB — requires 48GB VRAM or sequential offloading on 32GB. Place in models/checkpoints/.

Pair with the dev model (dev FP8 on 16 GB, BF16 dev on 32 GB+) and set the sampler to 8 steps, CFG = 1. At rank 384 / 7.6 GB it is the heaviest, highest-fidelity distillation LoRA — prefer it when you want maximum distilled quality and have the VRAM headroom.

Official Lightricks rank-384 distillation LoRA, v1.1. Applied on the dev model, it converts dev-style inference (many steps, CFG > 1) into distilled-style (8 steps, CFG = 1) — the official high-rank counterpart to Kijai's dynamic-rank LoRA. Place in models/loras/.

Previous v1.0 LoRA. Use v1.1 LoRA for latest quality.

Official distilled LoRA rank-384 v1.0. Pair with dev model. Place in models/loras/.

Use the official IC-LoRA ComfyUI workflow with LTXICLoRALoaderModelOnly + LTXAddVideoICLoRAGuide (Reference Downscale Factor 0.5).

Unified Canny + Depth control IC-LoRA. One LoRA covering edge and depth conditioning. Place in models/loras/.

For V2V motion transfer. Pair with the official IC-LoRA workflow at Reference Downscale Factor 0.5.

Motion-track IC-LoRA. Drive motion with sparse point trajectories (SpatialTrackerV2 or hand-drawn). Place in models/loras/.

Required: also download the HDR scene-emb file. Enables HDR text/image-to-video and SDR-to-HDR conversion.

16-bit HDR generation IC-LoRA. SDR→HDR via LogC3 transform. Place in models/loras/.

Required when using the HDR IC-LoRA.

Companion scene-embeddings file for the HDR IC-LoRA. Place in models/loras/.

Use for video lip-sync / dubbing. Pair with the audio VAE. Same file as the Dub-It repo — download one, not both.

Lip-dubbing IC-LoRA based on JustDubIt research. Joint audio-visual diffusion for dubbing. Place in models/loras/.

Lightricks' card recommends LoRA strength 1.4, 30 steps, guidance 4.0 at 768×448×121. Use the two-part prompt — 'Reference sheet: …' describing the panels, then 'Generated video: …' driving the action. Not a general T2V model: it expects a reference sheet.

Reference-sheet control IC-LoRA: conditions generation on a single composite image inventorying the characters, props, and location of a scene, so the generated video keeps those elements consistent. The most-downloaded LTX 2.3 Creative Lab adapter. Trained at one bucket only (768×448, 121 frames, 24 fps). Gated repo. Place in models/loras/.

Works with an empty or minimal prompt; if you prompt, describe only the masked/outpainted region, never the full scene. For inpainting, dilate the mask past the object and cover its shadows, reflections, and contact areas — the model can re-inpaint an object from a leftover shadow alone.

Mask-conditioned IC-LoRA covering two tasks: inpainting (fill or remove a masked region — object removal, background replacement, wardrobe change) and outpainting (extend the canvas horizontally or vertically). Conditions on the reference video plus a binary mask; unmasked area is left intact. Gated repo. Place in models/loras/.

Same weights as the LipDub repo (identical 2.47 GB file and hash) under the paper's newer 'Dub-It' name — download one, not both. Use the official dub-it workflow from the LTX-2 ComfyUI repository; pair with the LTX 2.3 audio VAE.

Lip-dubbing IC-LoRA from the JustDubIt paper (arXiv 2601.22143) — joint audio-visual diffusion that re-syncs a speaker's mouth to a new audio track. Conditions on video and audio at the same resolution as the output. Gated repo. Place in models/loras/.

Prompt with the DEBLUR trigger in Lightricks' two-part caption format. Start at strength 1.0 and lower toward 0.8 if you see haloing or ringing. Out of scope per the card: motion blur, compression artifacts, denoising, or upscaling already-sharp footage — extreme blur gets hallucinated, not reconstructed.

Restoration IC-LoRA that recovers sharpness from out-of-focus / defocused footage by conditioning on the blurry clip and regenerating it sharp, preserving subject, framing, and scene geometry. Whole-frame effect, no mask support. Trained at 960×544, 121 frames @ 24 fps. Gated repo. Place in models/loras/.

Prompt with the ENHANCE QUALITY trigger; published samples used strength 1.0. Frame count must satisfy (frames−1) % 8 == 0. It targets compression only — defocus blur, motion blur, grain, upscaling, and colorization are explicitly out of scope; use the Deblur or Colorization adapter for those.

Restoration IC-LoRA that removes compression artifacts — macroblocking, chroma bleed, ringing, banding — from low-bitrate video while keeping identity, framing, and geometry. Trained and validated at 960×544×121 @ 24 fps in both orientations. Gated repo. Place in models/loras/.

Prompt with the COLORIZE trigger; describe the same scene in both halves of the caption and change only the color language so the model alters nothing else. Strength 1.0. Not a deblur, denoise, decompression, or upscaling model.

Restoration IC-LoRA that adds natural color to grayscale, monochrome, or desaturated video while leaving identity, framing, and geometry untouched — only color changes. Trained and validated at 960×544×121 @ 24 fps in both orientations. Gated repo. Place in models/loras/.

Include the ADD WATER trigger in every prompt using the 'Reference shows … Edited shows … ADD WATER …' format. Trained on real water: other liquids (lava, slime, paint) generalize only loosely. Strengths ≥ 1.5 maximize drama but can warp faces and fine detail.

VFX IC-LoRA that adds water to a clip — rivers, surf, rain, waterfalls, floods, splashes, spray, wet-surface specularities — while keeping the subject's identity, clothing, pose, camera framing, and background geometry identical to the reference. Gated repo. Place in models/loras/.

Strength 1.0, 30 steps, guidance 3.0–4.0 — lower is brighter with more preserved detail, higher is darker with a stronger night look. Clips much longer than the ~4 s training length can drift toward the end; indoor and heavily artificial-light scenes are out of scope.

Relighting IC-LoRA that re-renders a daytime video as the same shot at night, preserving composition, framing, camera movement, and scene structure. The prompt steers the style of night (moonlight, color temperature, brightness); the reference dictates geometry and motion. Trained at 768×448 / 448×768, 97 frames @ 24 fps. Gated repo. Place in models/loras/.

Minimal prompt: 'relight the video to match the light-direction ball.' followed by one look phrase and a direction phrase (e.g. 'hard low-angle sunlight from the right') — do not describe the scene. Strength 1.0. Exteriors only; it relights, it does not restyle content.

Relighting IC-LoRA that relights an exterior video to a chosen sun direction and lighting hardness, driven by a small 'light-direction ball' composited into the reference video's corner. Trained at 1280×704 × 121 frames @ 24 fps. Gated repo. Place in models/loras/.

No trigger word — describe the desired empty scene rather than issuing a 'remove X' command; the training captions describe the clean result. Strength 1.0. Keep dimensions divisible by 32 and frame counts at 8k+1. Fails when the subject fills most of the frame (e.g. POV legs/torso) — use the In/Outpainting adapter with a mask there.

VFX IC-LoRA that removes people and other dynamic subjects from a source video and reconstructs the empty background — a clean plate of the same scene with framing and camera motion intact. Trained at 1024×576 / 576×1024, 49 frames @ 25 fps; validated at 1920×1088. Gated repo. Place in models/loras/.

Always start the prompt with the REMOVEBEARD trigger, then describe the clean-shaven subject and scene. Strength 1.0, 30 steps, guidance 4.0 with STG enabled. It will not grow or restyle hair; identity can drift on very long clips.

Portrait-edit IC-LoRA that removes facial hair (beard, mustache, stubble) from a person in a video while preserving identity, expression, motion, lighting, and framing. Trained at 960×544 / 544×960, 49 frames @ 25 fps; generalizes to ~121 frames. Gated repo. Place in models/loras/.

Needs a straight-eyed reference video and a frame count with frames % 8 == 1 (49, 145, 201…). Card example: 30 steps, guidance 4.0, STG scale 1.0 (stg_v), with a negative prompt of 'normal eyes, straight eyes, corrected eyes, gaze correction'.

Portrait-edit IC-LoRA that changes close-up portrait videos from straight eyes to convergent strabismus (inward-turned / crossed eyes) while preserving expression, head movement, lighting, and framing. Rank 32, 3000 training steps, buckets 960×544×49 and 544×960×49. Gated repo. Place in models/loras/.

Lower LoRA strength keeps output closer to the reference; higher allows more creative hallucination of detail. Not a denoiser or artifact remover, and not pixel-accurate — treat it as a creative step. The same repo ships a 4× variant for bigger jumps.

Creative pixel-space 2× upscaler IC-LoRA — synthesizes fine detail on a low-resolution video rather than interpolating it. Intended workflow: draft at very low base resolution (~280p) to lock composition and motion, then upscale. Gated repo. Place in models/loras/.

Choose 4× only when 2× is not enough of a jump; the further you scale, the more detail is hallucinated. Same strength trade-off as the 2× file.

4× variant of the creative pixel-space upscaler IC-LoRA — same repo and workflow as the 2× file, for larger resolution jumps from a low-res draft. Synthesizes new detail rather than preserving the reference. Gated repo. Place in models/loras/.

Put the CINEMAGRAPH_MOTION trigger in the positive prompt. Card recommends strength 0.7–3.0 and STG in stg_v mode, scale 1.0. Optimized for static-camera, single-element motion — camera moves, body motion, or multiple moving subjects are out of scope.

Style LoRA (not an IC-LoRA — no reference video) that turns a still image into a smooth looping cinemagraph: one element moves while the rest of the frame stays frozen, tripod-locked. Gated repo. Place in models/loras/.

Strength 0.8–1.0, adjust per clip. Write prompts as physical sound descriptions ending with 'No speech is present. No music is present.' Out of scope: dialogue, music, lip-sync, or a full production mix. Requires the LTX 2.3 audio VAE.

Video-to-audio Foley LoRA: generates synchronized, music-free sound effects from a silent video (also works for text-to-audio Foley prompts). Primary use is muting the source and letting the model produce matching SFX. Gated repo. Place in models/loras/.

Use this on 16/24 GB VRAM cards. Full BF16 Gemma OOMs alongside the transformer.

FP4-mixed Gemma 3 12B IT text encoder (~90% FP4 layers). Required for 16-24 GB ComfyUI workflows. Place in models/text_encoders/.

Alternative for 16-24 GB cards. Slightly higher quality than FP4 at 3.7 GB more.

FP8-scaled Gemma 3 12B IT text encoder. Alternative to FP4 with marginally higher precision. Place in models/text_encoders/.

Rename to comfy_gemma_3_12B_it.safetensors after downloading. Use on 32 GB+ cards for highest text-encoder quality.

Full BF16 Gemma 3 12B IT text encoder (HF filename gemma_3_12B_it.safetensors — workflows reference it as comfy_gemma_3_12B_it.safetensors, rename after download). Place in models/text_encoders/.

Download alongside your Gemma 3 12B text encoder — workflows that load the encoder as separate components also load this projection file. About 0.5 GB. Not needed if your workflow uses an all-in-one text encoder node.

Text projection layer (BF16) that maps the Gemma 3 text-encoder output into the dimension LTX 2.3's transformer expects. A small required component for ComfyUI workflows that wire the text encoder up as separate parts. Place in models/text_encoders/.

Previous version. Use v1.1 Distilled for latest quality.

v1.0 distilled version. 8 steps, CFG=1. Superseded by v1.1.

Previous version. Use v1.1 FP8 for latest quality.

FP8 distilled v3 by Kijai. Previous version, superseded by v1.1 FP8.

Previous version. Use FP8 v3 or v1.1 FP8 for better quality.

FP8 distilled v1 by Kijai. Earliest FP8 release, superseded by v3.

Previous version. Use FP8 v3 or v1.1 FP8 for better quality.

FP8 distilled v2 by Kijai. Superseded by v3.

Alternative FP8 quantization. Use v1.1 FP8 for latest quality.

FP8 scaled distilled by Kijai. Alternative FP8 quantization method.

Previous v1.0 MXFP8 (RTX 30xx). Use v1.1 MXFP8 for latest quality.

MXFP8 block-32 distilled v1.0 by Kijai. RTX 30xx workaround — use when standard FP8 scaled is unsupported.

Previous v1.0 BF16 distilled. Use v1.1 BF16 for latest quality. Enable Sequential Offloading on 32GB.

BF16 distilled v1.0 transformer-only by Kijai. 44GB — requires 48GB VRAM or sequential offloading on 32GB.

Prefer the dotted-filename version unless a specific workflow references this exact name.

FP8 input-scaled dev model under the alternate `ltx-2-3-…` (dash) filename. Same weights, different name.

Required for all setups. Download this first regardless of your VRAM.

VAE by Kijai. Required for all ComfyUI workflows. Place in models/vae/.

Required only for audio-conditioned / audio-to-video workflows. Pair it with the video VAE — the joint pipeline loads both. Not needed for silent T2V or I2V.

Audio VAE by Kijai for LTX 2.3's joint audio-video generation. Encodes/decodes the audio latent stream so a workflow can produce sound synchronized with the video. Place in models/vae/.

Most setups only need taeltx2_3.safetensors. Download this full BF16 VAE only if a workflow references it by name, or if you see banding/color drift with the tiny VAE on long clips.

Full-precision BF16 video VAE by Kijai — the heavier, higher-fidelity alternative to the tiny taeltx2_3 decoder. Use it when a workflow asks for it or when you need the cleanest decode on long or HDR clips. Place in models/vae/.

Worth it if VAE decode is your bottleneck — long clips, high resolution, or a 16GB card that OOMs at the decode step. Select it in the VAELoader node in place of LTX23_video_vae_bf16 or taeltx2_3; nothing else in the workflow changes.

ComfyUI-format conversion of Pruna AI's PrunaVAED — a drop-in replacement video VAE for LTX 2.3 with a pruned, faster decoder. Kijai's one-line summary: 'Faster decode, unchanged encode.' Pruna reports ~1.7× faster decoding and ~50% lower peak decode VRAM versus the stock BF16 video VAE, with the encoder and latent format unchanged, so existing workflows need no graph edits. Video VAE only — keep using LTX23_audio_vae_bf16.safetensors for audio. Place in models/vae/.

Optional. Generate at a lower base resolution for speed, then run this x2 latent upscaler as a second stage for a sharper final result. Prefer v1.1 over v1.0. Tiny (~1 GB) and light on VRAM.

Official v1.1 latent spatial upscaler — doubles the spatial resolution of LTX 2.3 output inside a two-stage pipeline, working in latent space rather than on decoded pixels. Place in models/latent_upscale_models/.

Optional. Use in a two-stage pipeline to upscale output resolution after generation.

Spatial upscaler x2 for two-stage pipelines. Place in models/latent_upscale_models/.

Use when x2 upscale is too aggressive. Gentler upscaling option.

Spatial upscaler x1.5 for two-stage pipelines. Place in models/latent_upscale_models/.

Optional. Use to double frame count (e.g. 65→129 frames) for smoother motion.

Temporal upscaler x2 for frame interpolation. Place in models/latent_upscale_models/.