Build Lightricks LTX-2 / LTX-2.3 video workflows covering text-to-video, image-to-video, GGUF and bundled checkpoints, distilled model, camera control LoRAs, synchronized audio, two-stage upscaling, and swapping alternate/GGUF base models
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Build Lightricks LTX-2 / LTX-2.3 video workflows covering text-to-video, image-to-video, GGUF and bundled checkpoints, distilled model, camera control LoRAs, synchronized audio, two-stage upscaling, and swapping alternate/GGUF base models
globs
["**/*.json"]
LTX-2 / LTX-2.3 Video Workflows
Version naming (read this first)
There is no "LTX 3.2" or "LTX2.3" as separate products. The user's shorthand refers to Lightricks LTX-2.3, a point release of the LTX-2 family. The lineage is:
LTX-Video (2024): first text-to-video model from Lightricks.
LTX-2 / LTX-V2 (Oct 2025): 19B-class DiT audio-video foundation model. Bundled checkpoint ltx-2-19b-distilled.safetensors, Gemma 3 12B text encoder.
LTX-2.3 (released ~March 2026): 22B-parameter DiT update. Rebuilt VAE (sharper textures/faces/hair/text), ~4x larger text connector (text projection) for prompt adherence, native 9:16 portrait, LoRA support, HiFi-GAN vocoder for cleaner synchronized audio, up to 4K@50fps / ~20s clips. Apache 2.0. Distributed primarily as GGUF UNets (community quants) plus separate VAE / text-encoder / text-projection files, NOT a single bundled checkpoint like LTX-2.
When the user says "LTX3.2" / "LTX2.3", treat it as LTX-2.3. This skill covers both LTX-2 (bundled checkpoint path) and LTX-2.3 (GGUF UNet path).
⭐ Render-verified correct setup (read this FIRST — 2026-06-19)
The GGUF-UNet + DualCLIPLoader + gemma_3_12B_it_fp4_mixed path documented later
in this skill (the Aitrepreneur installer path) It runs, but it is NOT the quality path. The setup
below is the official Comfy-Org template, render-proven sharp (1280×704, accurate
faces, synchronized 48 kHz stereo audio).
produces soft/mushy video with
inaccurate faces and eyes.
Models (exact, render-verified)
Component
File
Source repo
Folder
Notes
Checkpoint
ltx-2.3-22b-dev.safetensors (46 GB, max quality) orltx-2.3-22b-dev-fp8.safetensors (~23 GB, official VRAM-friendly)
Lightricks/LTX-2.3 / Lightricks/LTX-2.3-fp8
checkpoints/ (NOT unet/)
The checkpoint carries the transformer and the audio VAE. Loaded by CheckpointLoaderSimple + reused by LTXVAudioVAELoader + LTXAVTextEncoderLoader.
Gemma text encoder
gemma_3_12B_it_fp8_scaled.safetensors (13 GB)
Comfy-Org/ltx-2 → split_files/text_encoders/
text_encoders/
Use fp8_scaled (unpacked). The Aitrepreneur fp4_mixed mirror file is truncated (5.3 GB vs 9.4 GB) AND a packed-fp4 layout core can't reshape → shape [15360,1920] invalid for input 27582328.
Applied to the text-encoder CLIP via a LoraLoader. This is the prompt-accuracy / correct-eyes fix. Missing this = subtly-wrong faces.
Spatial upscaler
ltx-2.3-spatial-upscaler-x2-1.1.safetensors
Lightricks/LTX-2.3
latent_upscale_models/
Used by the stage-2 LTXVLatentUpsampler. Use x2-1.1, not x2-1.0.
Node stack (the right one)
LTXAVTextEncoderLoader (CORE, comfy_extras/nodes_lt_audio.py) loads gemma + the full checkpoint together via comfy.sd.load_clip([gemma, ckpt], type=LTXV). This is the audio-video encoder driving both video and audio/voice. Do NOT use DualCLIPLoader(type=ltxv) + a separate ltx-2.3_text_projection file. That is the legacy video-only path and yields mush.
Gemma abliterated LoRA via a LoraLoader (CLIP LoRA) on the encoder output → CLIPTextEncode.
Two-stage: base sample (~768×512) → LTXVLatentUpsampler (×2 spatial, uses the upscaler model + the checkpoint VAE) → refine sample → 1280×704 output. The upscale is the sharpness. A single-stage graph is visibly softer.
Guider: the Comfy-Org template uses plain CFGGuider cfg=1 (distilled); the LTXVideo repo example uses MultimodalGuider + GuiderParameters (separate AUDIO/VIDEO) + ClownSampler_Beta (RES4LYF). Both produce sharp output. The LoRAs + two-stage matter more than the guider.
ffmpeg is required for the final mux: <comfy-venv>/python -m pip install imageio-ffmpeg, then reboot. CreateVideo/SaveVideo/VHS_VideoCombine fail with ffmpeg ... could not be found otherwise.
Custom nodes
ComfyUI-LTXVideo (LTXV* nodes, MultimodalGuider, GuiderParameters, LTXVPreprocess, LTXVTiledVAEDecode, GemmaAPITextEncode, LTXFloatToInt) + RES4LYF (ClownSampler_Beta, only for the repo-example sampler). LTXAVTextEncoderLoader, ResizeImageMaskNode, CreateVideo, SaveVideo, ManualSigmas, LTXVScheduler, the Primitive* nodes are all CORE ComfyUI.
Quality troubleshooting (symptom → cause → fix)
Mushy/garbage, no clear subject → empty positive prompt, or DualCLIPLoader+projection text encoder. Fix: set a prompt; use LTXAVTextEncoderLoader.
Coherent but soft/blurry, faces & eyes slightly wrong → no two-stage upscale and/or missing the gemma abliterated LoRA and/or the old distilled LoRA. Fix: full two-stage template + both LoRAs above.
status: success but no video file / outputs only has a math or text node → the output node (SaveVideo/VHS) failed validation and was silently dropped; the graph short-circuited. Check the ComfyUI log for Failed to validate prompt for output N and fix that node (missing ffmpeg, a broken connection, a model-not-in-list).
DualCLIPLoader reshape [15360,1920] invalid for input 27582328 → wrong/truncated gemma → use gemma_3_12B_it_fp8_scaled.
LatentUpscaleModelLoader: ...x2-1.0 not in list → reference ...x2-1.1.
SaveVideo writes to a subfolder (video/<prefix>_NNNNN.mp4). Its history outputs entry isn't under images/videos/gifs, so a naive "find the video" check misses it. Look on disk under output/video/.
The official template exercised several convertUiToApi gaps, all now fixed. Keep them in mind if a template still mis-converts:
V3 dynamic combos (COMFY_DYNAMICCOMBO_V3, e.g. ResizeImageMaskNode.resize_type): each selected option's nested input must be keyed <combo>.<nested> (e.g. resize_type.longer_size, resize_type.width), NOT flat. ComfyUI rebuilds the nested dict via dynamic_paths/finalize_prefix. A flat key is rejected required_input_missing.
Reroute is virtual. Its connections must be passed through (consumer resolves to the Reroute's input), else everything downstream dangles and the graph short-circuits.
VHS_VideoCombine stores widgets_values as a name→value object, not a positional array.
Typed Primitive* nodes (PrimitiveInt/Float/Boolean/StringMultiline) are real executable nodes. Keep them as link sources; don't bake their values into a consumer's widgets_values by index (mis-positions V3 nested inputs).
Pack
packs/ltx-2.3-txt2vid (and the i2v/flf/extender variants) should be built on this official two-stage template. For a no-input-file T2V pack, set the template's bypass_i2v / "Switch to Text to Video?" boolean true and feed the I2V image input a blank EmptyImage (discarded at runtime but still validates).
Source note: the install scripts below pull LTX-2.3 files from a third-party mirror repo huggingface.co/Aitrepreneur/FLX, not the official Lightricks/LTX-2.3 repo. The official weights live at huggingface.co/Lightricks/LTX-2.3. Filenames/quants match what those scripts download.
Overview
LTX-2 is a DiT-based video foundation model from Lightricks. It uses a Gemma 3 12B text encoder and supports both text-to-video (T2V) and image-to-video (I2V). Key features:
Distilled model for fast 8-step generation; dev model for higher quality (~20+ steps)
Two-stage pipeline: Generate at low res, then 2x spatial upscale in latent space
Camera control LoRAs for cinematic movements
Synchronized audio-video generation in a single pass (LTX-2.3 audio VAE + HiFi-GAN vocoder)
GGUF quantization (LTX-2.3) for low-VRAM local inference via ComfyUI-GGUF
Loading note (LTX-2): The bundled checkpoint contains the VAE internally. The Gemma 3 text encoder loads separately via CLIPLoader with type: "ltxv" pointing at text_encoders/.
LTX-2.3 (GGUF UNet path — current install)
LTX-2.3 ships as a separate GGUF UNet + standalone VAE + text encoder + text projection, not a single bundled checkpoint. The install scripts (see below) place files like this:
Component
Node
Model file
Folder
Notes
UNet (GGUF)
UnetLoaderGGUF ("Unet Loader (GGUF)", bootleg category, from ComfyUI-GGUF)
22B dev model. Q4_K_S <12GB VRAM, Q5_K_S 12–16GB, Q8_0 24GB+
Video VAE
VAELoader
LTX23_video_vae_bf16.safetensors
models/vae/
rebuilt LTX-2.3 VAE
Audio VAE
VAELoader
LTX23_audio_vae_bf16.safetensors
models/vae/
only for audio-sync output
Gemma 3
CLIPLoader (type=ltxv)
gemma_3_12B_it_fp4_mixed.safetensors
models/text_encoders/
same FP4 encoder as LTX-2
Text projection
loaded with the text encoder
ltx-2.3_text_projection_bf16.safetensors
models/text_encoders/
the enlarged text connector new in 2.3
Spatial upscaler
LatentUpscaleModelLoader
ltx-2.3-spatial-upscaler-x2-1.1.safetensors
models/latent_upscale_models/
replaces LTX-2's ...x2-1.0
Loading note (LTX-2.3): Because the UNet is a bare GGUF, the VAE no longer comes "for free" with a checkpoint. Load LTX23_video_vae_bf16.safetensors explicitly with VAELoader. Place GGUF UNets in models/unet/ and use the GGUF Unet loader. Some community 2.3 workflows pair gemma_3_12B_it.safetensors (full) instead of the FP4 mixed file; the installer uses the FP4 mixed one.
Connect the optional latent input for latent-aware shift scaling.
Feeding a prior stage's output into I2V (e.g. Krea2 image → LTX video). The
LoadImage that feeds LTXVImgToVideo.image needs the source frame registered
as a ComfyUI INPUT. When that frame is an OUTPUT from an earlier stage, call
upload_image (action:"stage") with its { filename, subfolder?, type? } and drop
the returned input filename into LoadImage. (For a file already on local
disk, upload_image (action:"image").) NEVER copy the output file into, or guess, a
filesystem input/ path. ComfyUI's input/output dirs may be CUSTOM
(--input-directory / --output-directory), so a guessed path makes
LoadImage reject the file (Invalid image file) and wastes the render.
upload_image (action:"stage") goes through the server API (/view → /upload/image)
and resolves the real dirs correctly.
VERIFY A VIDEO RENDER VIA THE FILESYSTEM, NOT /history.VHS_VideoCombine
(and similar video nodes) write the .mp4 but frequently do NOT register the
output in ComfyUI's /history. The prompt shows done with an empty outputs map
and no error. Do NOT conclude the render "silently dropped" from
get_history / queue (action:"status") alone. Confirm the file with
get_image (action:"list_outputs") (it now lists videos too, with kind: "video"): match
the filename_prefix (e.g. ltxv2_….mp4), check the mtime is fresh, then
chain it into the next stage with upload_image (action:"stage").
LTXVImgToVideo (For I2V)
All-in-one node that encodes image, creates latent, and wraps conditioning:
Gotcha: strength controls motion; DON'T set it to 1.0.LTXVImgToVideo.strength
is how strongly the output adheres to the start image: higher = more adherence = LESS
motion. Setting it to 1.0 pins every frame to the start image → a FROZEN i2v with
ZERO motion (the storyboard frames come out nearly identical). Keep the verified
value ~0.6 (as in the example above) for proper motion. If a generated i2v clip
shows little/no motion, the FIRST thing to check is that strength wasn't bumped toward
1.0.
Requires LatentUpscaleModelLoader. Use ltx-2.3-spatial-upscaler-x2-1.1.safetensors for LTX-2.3 (or ltx-2-spatial-upscaler-x2-1.0.safetensors for LTX-2).
Sampler Settings
Distilled Model (Installed)
Uses SamplerCustomAdvanced with manual sigmas, NOT standard KSampler:
Concept/style LoRAs CAN be stacked with camera control LoRAs.
VRAM Considerations
Config
VRAM
Notes
bf16 checkpoint + FP4 Gemma
~24GB+
Tight on RTX 4090, may OOM
FP8 checkpoint + FP4 Gemma
~16-20GB
Recommended for 24GB GPUs
bf16 + tiled VAE decode
~22GB
Use LTXVSpatioTemporalTiledVAEDecode
VRAM warnings from MEMORY.md: "LTXV2 can OOM on 24GB — suggest FP8 quantized models or --lowvram"
Tips for 24GB GPUs
Use VAEDecodeTiled or LTXVSpatioTemporalTiledVAEDecode instead of standard VAEDecode
Start at 768x512 resolution, upscale in Stage 2
Use FP4 Gemma text encoder (installed)
For LTX-2.3, pick the GGUF quant to match VRAM: Q4_K_S (<12GB), Q5_K_S (12 to 16GB), Q8_0 (24GB+). The dev GGUF needs ~20+ steps; the distilled LoRA path runs ~8 steps
Always clear_vram before switching to LTX-V2 from another model family
Reduce frame count to 81 or 49 if OOM persists
Prompt Style
Natural language descriptions. Be specific about motion, camera angles, and temporal progression:
Good: "A woman with flowing auburn hair walks through a sun-dappled forest, leaves falling gently around her, soft golden hour lighting, cinematic depth of field"
Bad: "woman, forest, walking"
Describe the entire scene progression, not a single moment. Include lighting, mood, and motion cues.
Two-Stage Upscale Pattern
For production quality, generate at low resolution then upscale:
Stage 1: Generate at 768x512, 121 frames, 8 steps (distilled)
Stage 2: Resample the upscaled latent with 3-4 steps at CFG 1.0
Decode: Use tiled VAE decode for the larger resolution
This requires the spatial upscaler model in models/latent_upscale_models/: ltx-2.3-spatial-upscaler-x2-1.1.safetensors (LTX-2.3) or ltx-2-spatial-upscaler-x2-1.0.safetensors (LTX-2).
Using alternate / GGUF base models (incl. the "sulphur" model)
You can swap the LTX UNet for any LTX-2.3-compatible base model. The most-asked-about one is Sulphur 2 (the user's "sulphur2Base_dev.safetensors"; see name note below).
What Sulphur 2 actually is (verified June 2026)
It exists and is real. Sulphur 2 is an uncensored, realism-leaning finetune/derivative of LTX-2.3 (22B DiT), marketed as a drop-in replacement inside existing LTX-2.3 ComfyUI graphs (T2V + I2V + the other 2.3 formats). It is NOT its own architecture and is not LTX-2 (19B) compatible. It targets the LTX-2.3 stack (2.3 VAE + Gemma 3 text encoder + 2.3 text projection).
Filename caveat: there is no file literally named sulphur2Base_dev.safetensors. The real base checkpoints are sulphur_dev_bf16.safetensors (~46 GB) and sulphur_dev_fp8mixed.safetensors (~29 GB). There is also a distilled variant (sulphur_distil_bf16.safetensors) and a LoRA (sulphur_lora_rank_768.safetensors). Treat "sulphur2Base_dev" as the user's shorthand for the Sulphur 2 base dev checkpoint.
GGUF version: confirmed.vantagewithai/Sulphur-2-Base-GGUF hosts sulphur_dev-<quant>.gguf for Q3_K_S/M, Q4_0/1/K_S/K_M, Q5_0/1/K_S/K_M, Q6_K, Q8_0 (~10 to 23 GB). There is also a Civitai/Sulphur-2-distilled-fp8 and Civitai listings ("Sulphur 2 Base", "Rebels Sulphur 2 GGUF").
Hosting: HF SulphurAI/Sulphur-2-base (safetensors + a bundled Qwen-based prompt-enhancer GGUF), HF vantagewithai/Sulphur-2-Base-GGUF (the GGUF quants), and Civitai mirrors. Uncensored open weights are in scope to document. Nothing here is fabricated, but verify the exact repo/license yourself before downloading.
How to load it (it slots straight into the LTX-2.3 GGUF workflow above)
The GGUF quant is a different UNet and nothing more. Load it with the same UnetLoaderGGUF node and keep the rest of the 2.3 graph identical:
Put sulphur_dev-Q8_0.gguf (or your chosen quant) in models/unet/.
In the LTX-2.3 GGUF workflow above, change node "1":
Keep the same LTX-2.3 companions: VAELoader → LTX23_video_vae_bf16.safetensors, CLIPLoader (type=ltxv) → gemma_3_12B_it_fp4_mixed.safetensors, plus ltx-2.3_text_projection_bf16.safetensors. These must match the LTX-2.3 architecture. Do not pair it with LTX-2 (19B) VAE/encoder.
For the bf16/fp8 safetensors (non-GGUF) variants, load with the LTX checkpoint/diffusion-model loader the workflow uses for the safetensors path (Lightricks recommends the native LTX Video nodes documented at docs.ltx.video, not the auto-generated Diffusers snippet) rather than UnetLoaderGGUF.
Obey the same constraints as any LTX-2.3 gen: frame count 8n+1, resolution multiples of 32, LTXVConditioning frame_rate, dev model ~20+ steps / distilled ~8 steps.
General rule for ANY alternate LTX base model
To verify a third-party model is usable before wiring it up:
Confirm the architecture/version it was trained on (LTX-2 19B vs LTX-2.3 22B). Mixing a 2.3 UNet with a 2.0 VAE/encoder will fail or produce garbage.
For GGUF: requires the ComfyUI-GGUF custom node (installed by the scripts), file in models/unet/, loaded via UnetLoaderGGUF. Match the correct VAE + text encoder + text projection for that LTX version.
For safetensors finetunes: load like the matching official checkpoint, keep the official VAE/encoder of the same version.
If you only have a LoRA (e.g. sulphur_lora_rank_768.safetensors), apply it to the matching base UNet with LoraLoaderModelOnly instead of swapping the whole model.
Troubleshooting
LTXVideo "kornia" import error (pad ImportError)
Symptom: ComfyUI-LTXVideo fails to load with an ImportError from kornia.geometry.transform.pyramid because pad can no longer be imported. This happens with kornia 0.8.3+, which stopped exporting pad from that module.
What the fix does (FIX-LTXVIDEO-KORNIA.bat, run from the ComfyUI_windows_portable folder): it patches ComfyUI/custom_nodes/ComfyUI-LTXVideo/pyramid_blending.py:
Backs the file up to pyramid_blending.py.bak_kornia_fix.
Removes the broken pad, line from the from kornia.geometry.transform.pyramid import ( ... ) block.
Inserts a compatibility shim right after import torch.nn.functional as F:
# Compatibility fix for Kornia 0.8.3+ where pad is no longer exported here
pad = F.pad
Verifies pad = F.pad is present and the broken import is gone.
Manual equivalent if you don't run the .bat: edit pyramid_blending.py to delete pad, from the kornia import list and add pad = F.pad after the import torch.nn.functional as F line, then restart ComfyUI. (Alternatively, pin kornia to a pre-0.8.3 release, but the patch is the lighter-touch fix and is what the install set ships.)
LTXVideo version / workflow mismatch
The RunPod installer pins ComfyUI-LTXVideo to commit cd5d371518afb07d6b3641be8012f644f25269fc for workflow compatibility. If 2.3 workflows error on the latest LTXVideo, check out that commit. Torch is pinned to 2.4.0 + cu121; do not let a node's requirements.txt upgrade torch (the installers sanitize requirements to prevent this).
Sources
Official: none found.
Empirical: sampler values, wiring, and prompt notes from working graphs in packs/ and observed renders; not a vendor prompting guide.
