> ## Documentation Index
> Fetch the complete documentation index at: https://docs.comfy.org/llms.txt
> Use this file to discover all available pages before exploring further.
# LTX-2.5 DFR in ComfyUI: Image to Video with Audio
> Run the LTX-2.5 DFR workflow in ComfyUI: a three-stage Diffusion Fidelity Rendering pipeline that generates video with synchronized audio from one image.
Generate a video with synchronized audio from a single image through LTX-2.5's Diffusion Fidelity Rendering path. DFR is the production-quality recipe: a distilled transformer paired with a detailing IC-LoRA, with keyframes generated inside the clip and reused as guides by the later passes. Turn on `Switch to T2V` to run the same pipeline from text alone.
For the other LTX-2.5 workflows, see the [LTX-2.5 overview](/tutorials/video/ltx/ltx-2-5).
Open in Comfy Cloud
Download JSON or search "LTX-2.5 DFR" in Template Library
Download the default input image, or use your own image.
## The three stages
| Stage | Runs at | Steps | What it does |
| :- | :- | :- | :- |
| 1. Generate | Half the output size | 8 | Denoises the video, one detailing keyframe per second, and the audio together in one pass |
| 2. Spatial x2 | The output size | 3 | Upsamples the stage-1 latent and re-samples it with the pixel-spatial detailing IC-LoRA, using the half-resolution latent as reference |
| 3. Temporal x2 | 241 frames at 48 fps | 4 | Doubles the frame rate over the same duration, with the stage-2 keyframes pinned as guides |
Every stage samples with the `euler` sampler, `cfg` 1.0, and a fixed `ManualSigmas` schedule. The checkpoint is distilled, so the step counts and the schedules are part of the recipe: changing them changes the look of the result.
The third stage is optional. Turn off `Run 3rd stage - Temporal Upscale` to skip it and get a clip at the stage-2 frame rate instead. Audio is carried from stage 1 through the whole pipeline, so the sound does not change when the temporal pass runs.
## Model downloads
The DFR workflow loads the full-precision text encoder and adds the detailing IC-LoRA plus the temporal upscaler, so its model set differs from the other LTX-2.5 workflows.
Place in ComfyUI/models/diffusion\_models/
Place in ComfyUI/models/text\_encoders/
Optional: only needed when the prompt enhancer is enabled. Place in ComfyUI/models/text\_encoders/
Place in ComfyUI/models/loras/
Place in ComfyUI/models/latent\_upscale\_models/
Place in ComfyUI/models/latent\_upscale\_models/
Place in ComfyUI/models/vae/
Place in ComfyUI/models/vae/
## Model storage
```
ComfyUI/
├── 📂 models/
│ ├── 📂 diffusion_models/
│ │ └── ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors
│ ├── 📂 text_encoders/
│ │ ├── gemma4-12b-with-proj-ltx-2.5-bf16.safetensors
│ │ └── gemma4_e2b_it_bf16.safetensors (optional)
│ ├── 📂 loras/
│ │ └── ltx-2.5-22b-ic-lora-pixel-spatial-upscaler-x2-1.0.safetensors
│ ├── 📂 vae/
│ │ ├── ltx-2.5-video-vae-bf16.safetensors
│ │ └── ltx-2.5-audio-vae-bf16.safetensors
│ └── 📂 latent_upscale_models/
│ ├── ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors
│ └── ltx-2.5-latent-temporal-upscaler-x2-bf16-1.0.safetensors
```
## Workflow controls
* `Start image (I2V)`: the first frame. It is ignored while `Switch to T2V` is on
* `Output size (aspect / megapixels)`: the target aspect ratio and pixel budget, rounded to multiples of 32. Stage 1 runs at half of this size
* `Switch to T2V`: run the pipeline from text only, without a start image
* `Float (fps)` and `Float (Duration)`: 24 fps for 5 seconds by default, which is 121 frames
* `Enable Prompt Enhancer`: rewrite and extend the prompt with the small Gemma 4 encoder before sampling
* `Enable Tiled Decode`: decode the VAE in tiles to lower VRAM use, at a small cost in fine detail
## Prompting tips
1. **Write for one continuous shot**: DFR builds the clip around generated keyframes, so describe a scene that unfolds in a single take rather than a sequence of cuts
2. **Audio in the same pass**: put spoken lines in quotes and describe ambience and effects; the model generates the audio together with the video
3. **Keep camera moves gentle**: fast moves across noisy textures and smooth gradients are where the pipeline is most likely to show boiling or banding
This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.