NEW LTX-2 Workflows here: https://civarchive.com/models/2318870
Workflow: Image -> Autocaption (Prompt) -> LTX Image to Video
LTX Prompt Enhancer (LTXPE) might have issues with latest Comfy and Lightricks update
Update July 20th 2025: GGUF Models for LTX 0.9.8:
Distilled model, works with V9.5: https://huggingface.co/QuantStack/LTXV-13B-0.9.8-distilled-GGUF/tree/main
Dev model, works with V9.0: https://huggingface.co/QuantStack/LTXV-13B-0.9.8-dev-GGUF/tree/main
(see "Model Card" in above links for LTX 0.9.8 VAE and textencoder downloads)
V9.5: LTX 0.9.7 Distilled Workflow supporting LTX 0.9.7 Distilled GGUF Model.
There is a workflow with Florence and another one with LTX Prompt Enhancer (LTXPE)
GGUF Model can be downloaded here:
https://huggingface.co/wsbagnsv1/ltxv-13b-0.9.7-distilled-GGUF/tree/main
VAE and Textencoder are identical to previous LTX 0.9.6 model (see V8.0 below)
LTX 0.9.7 Distilled is using only 8 steps and is very fast.
V9.0: LTX 0.9.7 Workflow supporting LTX 0.9.7 GGUF Model.
There is a workflow with Florence and another one with LTX Prompt Enhancer (LTXPE)
GGUF Model can be downloaded here:
https://huggingface.co/wsbagnsv1/ltxv-13b-0.9.7-dev-GGUF/tree/main
VAE and Textencoder are identical to previous LTX 0.9.6 model (see V8.0 below)
LTX 0.9.7 is a 13billion parameter model, previous versions only had 2b parameters, therefore it is more heavy on Vram usage and requires longer process time. Try V8.0 below with model 0.9.6 or V9.5 for very fast rendering.
V8.0: LTX 0.9.6 Workflow (dev and distilled GGUF model in same workflow)
there is a version with Florence2 Caption and a version with LTX Prompt Enhancer (LTXPE)
GGUF Models (Dev & Distilled) can be downloaded here:
https://huggingface.co/calcuis/ltxv0.9.6-gguf/tree/main
vae: pig_video_enhanced_vae_fp32-f16.gguf
Textencoder: t5xxl_fp32-q4_0.gguf
V7.0: LTX 0.9.5 Model Version GGUF with Wavespeed/Teacache.
LTX 0.9.5 GGUF Model and VAE: https://huggingface.co/calcuis/ltxv-gguf/tree/main
(vae_ltxv0.9.5_fp8_e4m3fn.safetensors)
Clip Textencoder: https://huggingface.co/city96/t5-v1_1-xxl-encoder-gguf/tree/main
There are 2 worklfows, a main workflow with florence caption only and additional one with florence and LTX prompt enhancer. Setup with Wavespeed (bypassed by default, Strg+B to activate)
workflow works with all GGUF models: 0.9 / 0.9.1 / 0.9.5
uncensored LLM for Prompt enhancer: https://huggingface.co/skshmjn/unsloth_llama-3.2-3B-instruct-uncenssored
-Outdated (march 2025)- V6.0: GGUF/TiledVAE Version & Masked Motion Blur Version
Updated the workflow with GGUF Models, which save Vram and run faster.
There is a Standard Version, which uses just the GGUF Models and a GGUF+TiledVae+Clear Vram Version, that reduces Vram requirements even further. Tested the larger GGUF model (Q8) with resolution of 1024, 161 frames and 32 steps , the GGUF Version peaked Vram usage at 14gb, while the TiledVae+ClearVram Version peaked at 7gb. Smaller GGUF Models might reduce requirements further.
GGUF Model, VAE and Textencoder can be downloaded here:
(Model&VAE): https://huggingface.co/calcuis/ltxv-gguf/tree/main
(anti Checkerboard Vae): https://huggingface.co/spacepxl/ltx-video-0.9-vae-finetune/tree/main
(Clip Textencoder): https://huggingface.co/city96/t5-v1_1-xxl-encoder-gguf/tree/main
You can go for the GGUF Version with 16gb+ and the TiledVae+ClearVram with less than 16gb Vram.
Masked Motion Blur Version: Since LTX is prone to motion blur, added an extra group to the workflow which allows to set a mask on input image, apply motion blur to mask, to trigger specific motion. (sounds better than it actually works, useful tho in some cases). GGUF and GGUF+TiledVAE+ClearVram version included.
V5.0: Support for new LTX Model 0.9.1.
included an additional workflow for LowVram (Clears Vram before VAE)
added a workflow to compare LTX Model 0.9.1 vs LTX Model 0.9
(V4 did not work with 0.9.1 when the model was released (hence v5 was created), this has changed as comfy & nodes were updated in the meantime, now you can use both Models (0.9 & 0.9.1) with V4, also with V5. Both have different custom nodes to manage the model, other than that, both versions are the same. If you run into memory issues/long process time, see tips at the end)
-Outdated (march 2025)- V4.0: Introducing Video/Clip extension :
Extend a clip based on last frame from previous clip. You can extend a clip about 2-3 times before quality starts to degenerate, see more details in the notes of the worflow.
Added a feature to use your own prompt and bypass florence caption.
V3.0: Introducing STG (Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling).
Included a SIMPLE and an ENHANCED workflow. Enhanced Version has additional features to upscale the Input Image, that can help in some cases. Recommend to use the SIMPLE Version.
replaced the height/width Node with a "Dimension" node that drives the Videosize (default = 768. increase to 1024 will improve resolution, but might reduce motion, also uses more VRAM and time). Unlike previous Versions, Image will not be cropped.
Included new node "LTX Apply Perturbed Attention" representing the STG settings (for more details on values/limits see the note within the workflow) .
Enhanced Version has an additional switch to upscale Input Image (true) or not (false). Plus a scale value (use 1 or 2) to define the size of the image before being injected, which can work a bit like supersampling. As said, not required in most cases.
Pro Tip: Beside using the CRF value at around 24 to drive movement, increase the frame rate in the yellow Video Combine node from 1 to 4+ to trigger further motion when outcome is too static.
Node "Modify LTX Model" will change the model within a session, if you switch to another worklfow, make sure to hit "Free model and node cache" in comfyui to avoid interferences. If you bypass this node (strg-B) , you can do Text2Video.
V2.0 ComfyUI Workflow for Image-to-Video with Florence2 Autocaption (v2.0)
This updated workflow integrates Florence2 for autocaptioning, replacing BLIP from version 1.0, and includes improved controls for tailoring prompts towards video-specific outputs.
New Features in v2.0
Florence2 Node Integration
Caption Customization
A new text node allows replacing terms like "photo" or "image" in captions with "video" to align prompts more closely with video generation.
V1.0: Enhanced Motion with Compression
To mitigate "no-motion" artifacts in the LTX Video model:
Pass input images through FFmpeg using H.264 compression with a CRF of 20–30.
This step introduces subtle artifacts, helping the model latch onto the input as video-like content.
CRF values can be adjusted in the yellow "Video Combine" node (lower-left GUI).
Higher values (25–30) increase motion effects; lower values (~20) retain more visual fidelity.
Autocaption Enhancement
Text nodes for Pre-Text and After-Text allow manual additions to captions.
Use these to describe desired effects, such as camera movements.
Adjustable Input Settings
Width/Height & Scale: Define image resolution for the sampler (e.g., 768×512). A scale factor of 2 enables supersampling for higher-quality outputs. Use a scale value of 1 or 2. (changed to dimension node in V3)
Pro Tips
Motion Optimization: If outputs feel static, incrementally increase the CRF & frame rate value or adjust Pre-/After-Text nodes to emphasize motion-related prompts.
Fine-Tuning Captions: Experiment with Florence2’s caption detail levels for nuanced video prompts.
If you run into memory issues (OOM or extreme process time) try the following:
use the LowVram version of V5
use a GGUF Version
press "free model and node cache" in comfyui
set starting arguments for comfyui to --lowvram --disable-smart-memory
see the file in your comfyui folder: "run_nvidia_gpu.bat" edit the line: python.exe -s ComfyUI\main.py --lowvram --disable-smart-memory
switch off hardware acceleration in your browser
Credits go to Lightricks for their incredible model and nodes:
Description
Introducing STG (Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling).
FAQ
Comments (21)
Congratulations OP. With V3.0, you have officially provided us with true LTX I2V that uses STG. Even though most of the work was done by the node and model creators, you have used your extensive knowledge and gone out of the way to provide us less educated easy access to an LTX I2V workflow that uses STG. From what I can tell, this is the first widely accessible workflow released that uses STG and produces insane results. With 12gb of VRAM, I can finally input a 1024x1024 image and get an insane quality output of boobies bouncing in less than 20 seconds. For that, I give you all Buzz I have to offer. Congratulations again my good sir. Keep up the good work. O.O
Dude has sent me all his buzz, that is nuts and very kind of you, thank you. Glad this workflow works for you.
Let me know in case you are now out of buzz and broke,I am happy to return it.
yoo is the 1024x1024 getting you good results. do you know what the ideal resolution is or is it hit or miss
@iceburn 768 would say triggers more motion, 1024 sometimes generates static results, but also depends on the type of the input image.
@tremolo28 768 soulds really good, so it can be upscaled later. Can it do anime live2d animations?
any tips on what works best for pre/after text prompts? thank you for sharing this. best thing i've seen so far that will run on 12gb vram.
For most of the time I leave the pre/after text empty. if nothing moves, I add „it moves“. Sometimes I use it to force a camera movement, like the clip with the zoom into a hole in the wall with a house behind.
Thanks Man This work is a great breakthrough and a wonderful effort. Thank you.
works like a charm!
What are the major settings/parameters to control and specify the motion more precisely?
CRF? Frame Rate? Pre/After Text?
I tried different values, however either nothing moves or it just does random stuff...
e.g. "it moves" either pre/after text results in their mouth moving but the rest of the image just being static - or the whole image gets messed up, like arms deformed, fading to black etc.
Any tips?
I start with CRF = 20-22 and Frame Rate =1, if there is low motion I increase CRF up to 24-26, then increase Frame Rate until there is sufficient motion (usually around 4). If motion is too fancy, I reduce the values. It depends on the input image, different images can have different values. If you increase dimension (=resolution) from 768 to 1024, the CRF/Framrate might need a notch up to trigger motion. The great thing is, the model is very fast, so for me no issue if it takes 2-3 generations until you push it in the right direction.
Amazing work. Was getting decent results with LTX before, but this adds much more movement in the right context and areas. For the record I'm using 8gb RTX 4060, 2 min generation time on default settings (adjusted the CRF/Framerate as suggested). 3.96 it/s. Performing nicely on minimal VRam.
Hey thanks, good to know it works with 8gb Vram as well
How to use other llm other than Florence to do the captioning? It is fairly censored, even for PG stuff. Other than that, magnificent work!
You could replace the florence part in the workflow with any other captioning model/workflow.
@tremolo28 Comfy dummy here - how can I let florence do the initial caption, then let me edit what florence did before sending it to the next node?
@HarryPsalms Check the green node "CLIP Text Encode (Positive Prompt)" in the florence section of the workflow and right-click it and select "convert Input to Widget", then a textfield will apear to be used for prompting. You will still see the florence caption in the gui. you can copy the florence prompt into the textfield and edit it as you need it.
@tremolo28 When I do this it doesn't seem to do anything and the text field itself is populated by some default text.
I'm using Meta-Llama-3.1-8B for captioning, its better for nsfw.
using this custom node (just follow the instructions):
https://github.com/EvilBT/ComfyUI_SLK_joy_caption_two/blob/main/readme_us.md
@FluxFest check the experimental workflow, it includes an option to use your own prompt.
added an Ollama Version to Experimental Tab.