Uploaded August 2026 | Updated November 2026, 4 days ago
I tested LTX 2.5 in ComfyUI with action, physics, multishot, dialogue, Image-to-Video, GGUF and frame-control workflows.
You will see what LTX 2.5 improves, where it still breaks, how it compares with MiniMax H3 in difficult action scenes, and the workflows I used for the tests.
I start with harder action scenes because physics and destruction have always been difficult for LTX. The water, environment and synchronized audio look much better in LTX 2.5, but complex impacts and destruction can still feel unnatural.
I also compare some of these scenes with MiniMax H3. In my destruction tests, MiniMax H3 handled the impact and physics better, while LTX 2.5 still showed clear improvements over LTX 2.3.
Then I move into the ComfyUI workflows.
In this video I test:
• LTX 2.5 Text-to-Video
• Image-to-Video
• Native multishot generation
• Prompt Enhancer
• LTXV Duration Predictor
• Q4 GGUF workflow
• Distilled INT8 workflow
• Dialogue and character interaction
• Motion and vehicle physics
• First Frame + Last Frame
• First + Middle + Last Frame
• LTX 2.5 vs MiniMax H3 action tests
For one simple Text-to-Video test on my system, the generation took about 8 seconds.
I also test the Duration Predictor instead of manually guessing the video length. In one multishot example, I manually used 5 seconds, while the predictor estimated around 6 seconds. In the longer dialogue test, my manual setting was 15 seconds, while the predictor estimated around 19 seconds.
For the GGUF test, I use a Q4 model. It works, but in my setup I still prefer the distilled INT8 version when enough memory is available because I get a better result and the model-size difference is only around 5 GB.
I also test an older LTX 2.3 LoRA at 0.5 strength. That specific LoRA improved the physics in one of my LTX 2.5 tests, although the result was still not perfect.
The main reason I wanted to make this video was to show the real results. LTX 2.5 is clearly improved and very fast, but complex destruction, physics and difficult interactions can still fail.
RESOURCES
My LTX 2.5 ComfyUI workflows:
aistudynow.com/ltx-2-5-in-comfyui-what-improved-best-workflow-settings-and-my-tests
Official LTX 2.5 model:
huggingface.co/Lightricks/LTX-2.5
Official ComfyUI-LTXVideo nodes and workflows:
github.com/Lightricks/ComfyUI-LTXVideo
Official LTX documentation:
docs.ltx.io/open-source-model/getting-started/overview
Which LTX 2.5 test should I try next: physics, dialogue, multishot, longer videos or another workflow?
#LTX25 #ComfyUI #AIVideoGeneration
I tested LTX 2.5 in ComfyUI with action, physics, multishot, dialogue, Image-to-Video, GGUF and frame-control workflows.
You will see what LTX 2.5 improves, where it still breaks, how it compares with MiniMax H3 in difficult action scenes, and the workflows I used for the tests.
I start with harder action scenes because physics and destruction have always been difficult for LTX. The water, environment and synchronized audio look much better in LTX 2.5, but complex impacts and destruction can still feel unnatural.
I also compare some of these scenes with MiniMax H3. In my destruction tests, MiniMax H3 handled the impact and physics better, while LTX 2.5 still showed clear improvements over LTX 2.3.
Then I move into the ComfyUI workflows.
In this video I test:
• LTX 2.5 Text-to-Video
• Image-to-Video
• Native multishot generation
• Prompt Enhancer
• LTXV Duration Predictor
• Q4 GGUF workflow
• Distilled INT8 workflow
• Dialogue and character interaction
• Motion and vehicle physics
• First Frame + Last Frame
• First + Middle + Last Frame
• LTX 2.5 vs MiniMax H3 action tests
For one simple Text-to-Video test on my system, the generation took about 8 seconds.
I also test the Duration Predictor instead of manually guessing the video length. In one multishot example, I manually used 5 seconds, while the predictor estimated around 6 seconds. In the longer dialogue test, my manual setting was 15 seconds, while the predictor estimated around 19 seconds.
For the GGUF test, I use a Q4 model. It works, but in my setup I still prefer the distilled INT8 version when enough memory is available because I get a better result and the model-size difference is only around 5 GB.
I also test an older LTX 2.3 LoRA at 0.5 strength. That specific LoRA improved the physics in one of my LTX 2.5 tests, although the result was still not perfect.
The main reason I wanted to make this video was to show the real results. LTX 2.5 is clearly improved and very fast, but complex destruction, physics and difficult interactions can still fail.
RESOURCES
My LTX 2.5 ComfyUI workflows:
aistudynow.com/ltx-2-5-in-comfyui-what-improved-best-workflow-settings-and-my-tests
Official LTX 2.5 model:
huggingface.co/Lightricks/LTX-2.5
Official ComfyUI-LTXVideo nodes and workflows:
github.com/Lightricks/ComfyUI-LTXVideo
Official LTX documentation:
docs.ltx.io/open-source-model/getting-started/overview
Which LTX 2.5 test should I try next: physics, dialogue, multishot, longer videos or another workflow?
#LTX25 #ComfyUI #AIVideoGeneration










![USO (New IP-Adapter Flux) ComfyUI Workflow: Image Style Transfer, Same Face, Full Demos
Quick demo of the USO Lora (New IP-Adapter Flux) by bytedance research ComfyUI Workflow with FLUX where I copy style from a reference image and keep the same face across looks]
In this video, youll learn:
• how I set up the files and the LoRA so style transfer from image just works
• how I lock the subject face while swapping styles in ComfyUI
• what guidance settings and sizes gave me fast, clean results
Here’s what actually happened. I wired the USO ComfyUI Workflow, dropped in a style image, and the model picked up the look right away. Same prompt, new vibe. Then I added a subject image so the face stays the same when I move from manga lines to a wool texture or a Ghibli look. Let’s be real, I didn’t expect this part to be so quick. With FLUX and the USO LoRA, it worked better than I thought. If you’re on smaller VRAM, keep the sizes modest and it still feels snappy. I’ll show the exact nodes I used, the guidance band that behaved for me, and a couple of easy fixes when the style looks weak. Anyway, if you’re searching for a solid FLUX ComfyUI workflow, this should save you some testing time.
Resources mentioned:
• Free USO ComfyUI workflow: https://aistudynow.com/uso-comfyui-workflow-for-flux-style-from-image-same-face/
#USOComfyUI #FLUX #StyleTransfer USO (New IP-Adapter Flux) ComfyUI Workflow: Image Style Transfer, Same Face, Full Demos](https://i.ytimg.com/vi/_FrdkX0ReR8/mqdefault.jpg)