## ⚠️ Installation — Download Both Files
This workflow requires both downloadable files:
1. Config JSON — the ComfyUI workflow
2. Additional Components ZIP — required ltx25_smart_controls custom nodes
Select and download both files from the Download panel.
Extract the Additional Components ZIP directly into your ComfyUI root folder:
ComfyUI/custom_nodes/ltx25_smart_controls/
Then restart ComfyUI and import the workflow JSON.
Model weights are not included and must be installed separately.
## LTX-2.5 INT8 AllModes Workflow
A two-stage ComfyUI workflow for LTX-2.5 using the official Distilled INT8 ConvRot transformer.
### Features
- T2V, I2V and FLF2V in one workflow
- Final resolution, FPS, duration, exact frame count and seed controls
- LTX audio generation
- Optional standard LTX transformer LoRA
- Conv VAE tiled decoding
- Stage 1: SageAttention → Sol-Attn
- Selectable Stage 2 mode:
Speed — Sage + Sol
Designed for lower-VRAM systems and long clips.
12–16GB is the target range, but 12GB has not yet been physically verified.
Quality — Native Exact Attention
Uses native dense attention for the final refinement stage.
24GB+ VRAM is recommended, while 32GB is safer for 1920×1088 / 15-second generation.
### Tested Environment
- Windows
- NVIDIA RTX 5080 16GB
- 1920×1088 at 24 FPS
- 10-second Quality mode completes normally
- 15-second Quality mode completes but Stage 2 becomes significantly slower
- Speed mode is intended to reduce this full-resolution attention bottleneck
Stage 2 uses deterministic Euler with:
0.909375, 0.725, 0.421875, 0.0
### Required Custom Nodes
- ltx25_smart_controls v1.2.0 — must be installed separately
- ComfyUI-KJNodes 1.5.0
- ComfyUI-SolAttn_triton commit 842c4eaa7d91dbaef3fee3ccdbf36a39521e82fc
- SageAttention
- Triton / Triton-Windows
### Required Models
- ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors
- gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors
- ltx-2.5-video-vae-conv-bf16.safetensors
- ltx-2.5-audio-vae-bf16.safetensors
- ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors
Model weights are not included.
After importing the workflow, select your own FIRST image for I2V or FIRST/LAST images for FLF2V.
Built with reference to the official Lightricks LTX-2.5 two-stage workflow:
https://github.com/Lightricks/ComfyUI-LTXVideo
This is an experimental community optimization. Performance and VRAM requirements vary by resolution, duration, GPU, drivers and installed kernels.
Description
## ⚠️ Installation — Download Both Files
This workflow requires both downloadable files:
1. Config JSON — the ComfyUI workflow
2. Additional Components ZIP — required ltx25_smart_controls custom nodes
Select and download both files from the Download panel.
Extract the Additional Components ZIP directly into your ComfyUI root folder:
ComfyUI/custom_nodes/ltx25_smart_controls/
Then restart ComfyUI and import the workflow JSON.
Model weights are not included and must be installed separately.
Initial public release of an independent community workflow for LTXV 2.5.
### Features
- T2V, I2V, and FLF2V in one organized workflow
- Official LTX-2.5 Distilled INT8 ConvRot transformer
- Two-stage generation with the official x2 latent upscaler
- Stage 1: SageAttention → Sol-Attn
- Stage 2 Speed: SageAttention + Sol-Attn
- Stage 2 Quality: native exact attention
- Optional standard LTX transformer LoRA
- Video and LTX audio output
### Stage 2 modes
Speed — Sage + Sol
Designed for lower-VRAM systems and longer clips.
12–16GB is the target range, but a physical 12GB GPU has not yet been verified.
Quality — Native Exact
24GB+ VRAM is recommended. For 1920×1088 and 15-second clips, 32GB is safer.
### Tested environment
- Windows 11
- NVIDIA GeForce RTX 5080 16GB
- 1920×1088 at 24 FPS
- 10-second Quality mode tested successfully
This is a ComfyUI workflow, not a trained model. Training epochs and steps are not applicable.
Model weights and LoRA files are not included. Users must select their own FIRST image for I2V and FIRST/LAST images for FLF2V.
### Required custom nodes
- ltx25_smart_controls v1.2.0
- ComfyUI-KJNodes 1.5.0
- ComfyUI-SolAttn_triton commit 842c4eaa7d91dbaef3fee3ccdbf36a39521e82fc
- SageAttention
- Triton or Triton-Windows
Developed with reference to and incorporating elements of the official Lightricks ComfyUI-LTXVideo LTX-2.5 two-stage distilled workflow. This is a modified community workflow and is not endorsed by Lightricks or ComfyUI. Upstream and dependency licenses continue to apply.
FAQ
Comments (7)
i've used this workflow. it work good on rtx 3060 12gb and 32gb ram. i've generated 11 second videos in 630.22s (10.30 minutes). it's not the fastest but ok for me. thank you for this workflow
Thank you for sharing your feedback! I'm glad to hear it works well on the RTX 3060 12GB. ~10 minutes for an 11-second video is a pretty solid result for this setup. Enjoy generating!
Hi works fast on rtx 4070 12 gb and 32 gb ram on ram. i714700kf proccesor. It generates 10 seconds in 704x1280 resolution in about 140-160 seconds. I wanted to know if there is any chance to control the sampler steps on any stage to increase quality? , will upload some videos after finish testing this out
How can i change the number of the steps?
I get this error:sageattention is not new enough version or could not determine CUDA architecture, cannot apply LTX2 Memory Efficient Sage Attention Patch.
I installed sage and torch on my SwarmUI install as per the swarmUI docs, I dont know how to check the version or what version I even need.
Whenever I try and generate anything outside of the workflow with safe enabled via console I also get this:
//?/G:/SwarmUI/Data/tmp/82092/tmpb61pb9wi/cuda_utils.c:18: error: include file 'Python.h' not found
23:15:24.856 [Warning] [ComfyUI-0/STDERR] ←[1m←[31m[ERROR]←[0m Error running sage attention: Command '['G:\\SwarmUI\\dlbackend\\comfy\\python_embeded\\Lib\\site-packages\\triton\\runtime\\tcc\\tcc.exe', '\\\\?\\G:\\SwarmUI\\Data\\tmp\\82092\\tmpb61pb9wi\\cuda_utils.c', '-O3', '-shared', '-Wno-psabi', '-o', '\\\\?\\G:\\SwarmUI\\Data\\tmp\\82092\\tmpb61pb9wi\\cuda_utils.cp313-win_amd64.pyd', '-fPIC', '-D_Py_USE_GCC_BUILTIN_ATOMICS', '-lcuda', '-lpython313', '-LG:\\SwarmUI\\dlbackend\\comfy\\python_embeded\\Lib\\site-packages\\triton\\backends\\nvidia\\lib', '-LG:\\SwarmUI\\dlbackend\\comfy\\python_embeded\\Lib\\site-packages\\triton\\backends\\nvidia\\lib\\x64', '-IG:\\SwarmUI\\dlbackend\\comfy\\python_embeded\\Lib\\site-packages\\triton\\backends\\nvidia\\include', '-IG:\\SwarmUI\\dlbackend\\comfy\\python_embeded\\Lib\\site-packages\\triton\\backends\\nvidia\\include', '-I\\\\?\\G:\\SwarmUI\\Data\\tmp\\82092\\tmpb61pb9wi', '-IG:\\SwarmUI\\dlbackend\\comfy\\python_embeded\\Include']' returned non-zero exit status 1., using pytorch attention instead.
Which means something with sage is broken and IDK what causes it.
What sol attention doing here? Its for bf16 only. With int8 it gives only slowdown.
ltx25_smart_controls not found, google cant find it.
If you use exclusive not public nodes - share it with workflow or your workflow useless completely.