Skip to navigation
Reference

Workflow & Asset Compatibility

Use this page before assembling or modifying a workflow. Choose the row that matches your input and goal, then follow its linked workflow or template.

This is a reference guide, not an exhaustive listing of every workflow JSON in the repository.

Unless a row explicitly names another surface, these are ComfyUI workflows. For other native Python capabilities, see the PyTorch API; for the hosted API, see the API documentation.

Compatibility rules

  • Keep LTX-2.5 and LTX-2.3 workflows and assets separate unless a workflow or model card explicitly establishes compatibility.
  • Start from an official template or workflow, run it unchanged once, then add adapters or custom nodes one at a time.
  • An adapter’s model card is the source of truth for its compatible base model, loader, prompt wording, and strength.
  • Standard LoRAs and IC-LoRAs are different paths. The official LTX-2.5 workflow directory has task-specific IC-LoRA workflows. A standard LoRA requires manually adding ComfyUI’s core LoraLoaderModelOnly node to the selected baseline template.
  • Do not infer native ltx-pipelines file paths or supported features from a ComfyUI workflow (or the reverse).

Standard ComfyUI model files (BF16)

Use this higher-quality BF16 configuration as the reference stack for a standard LTX-2.5 ComfyUI workflow. A workflow can omit parts of it: for example, Text-to-Audio has no video branch and a single-stage workflow does not use the spatial upscaler.

FileRoleComfyUI folderUsed by
ltx-2.5-22b-distilled-transformer-bf16.safetensorsDistilled LTX-2.5 transformerComfyUI/models/diffusion_models/Standard LTX-2.5 workflows in this reference
gemma4-12b-with-proj-ltx-2.5-bf16.safetensorsLTX text encoderComfyUI/models/text_encoders/Prompt-conditioned LTX-2.5 workflows
gemma4_e2b_it_bf16.safetensorsOptional local prompt enhancerComfyUI/models/text_encoders/Official local templates that enable Prompt Enhance
ltx-2.5-video-vae-bf16.safetensorsVideo VAEComfyUI/models/vae/Video workflows; not Text-to-Audio
ltx-2.5-audio-vae-bf16.safetensorsAudio VAEComfyUI/models/vae/Workflows that generate or condition on audio
ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors2× spatial upscalerComfyUI/models/latent_upscale_models/Two-stage workflows only

Lower-VRAM ComfyUI configuration

For a lower-VRAM local setup with the distilled LTX-2.5 video workflows, use these three files together instead of their BF16 counterparts in the reference stack.

FileRoleComfyUI folder
ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensorsINT8 ConvRot distilled transformerComfyUI/models/diffusion_models/
gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensorsINT8 ConvRot Gemma text encoderComfyUI/models/text_encoders/
ltx-2.5-video-vae-conv-bf16.safetensorsConv video VAE; faster and lighter than the DiffVAEComfyUI/models/vae/

Text-to-Audio has no video branch, so it uses only the INT8 ConvRot transformer and text encoder together with the BF16 audio VAE.

The *-comfy-int8-convrot files are ComfyUI-only. Do not use them with native ltx-pipelines or the PyTorch API; use the BF16 files for those paths instead.

LTX-2.5 core workflows

GoalTemplateRequired inputKey compatibility notes
Text-to-VideoComfyUI template video_ltx2_5_t2vText promptTwo-stage template; generates synchronized video and audio. Start with the reference BF16 stack, or use the lower-VRAM configuration above.
Image-to-VideoComfyUI template video_ltx2_5_i2vSource image and supporting promptTwo-stage template. Use First-Frame / Last-Frame instead when both endpoints are supplied.
First-Frame / Last-FrameComfyUI template video_ltx2_5_flf2vStart image and end imageSingle-stage template for motion between two frames; do not assume the two-stage upscaler is used.
Audio-to-VideoLTX-2.5_A2V_Two_Stage_Distilled.jsonInput audio and a visual prompt; first image is optionalTwo-stage. The audio is the primary conditioning signal and is held fixed while video is generated.
Text-to-AudioLTX-2.5_T2A_Single_Stage_Distilled.jsonText promptAudio-only: use the audio VAE, not the video VAE or spatial upscaler.
Advanced two-stage T2V/I2VLTX-2.5_T2V_I2V_Two_Stage_Distilled.jsonText prompt; source image when using I2VBase generation followed by 2× spatial upscaling and refinement. Use the LTX-2.5 spatial upscaler.

LTX-2.5 IC-LoRA workflows

IC-LoRAs condition a generation on a reference input or control signal. Choose the task-specific workflow below, then follow its adapter card for the required assets, prompt wording, loader, and strength.

WorkflowWhat you need to know
Generic video-to-video IC-LoRA workflowInput: Reference video and a compatible adapter. Notes: Single-stage; follow the adapter card for its checkpoint, loader, prompt, and strength.
Union Control workflowInput: Reference video plus depth, canny, or pose control. Notes: ComfyUI-VideoDepthAnything and comfyui_controlnet_aux for the documented preprocessors.
Motion Control workflowInput: Prompt, opening image, and drawn sparse motion tracks. Notes: The opening image conditions the first frame and is the canvas for the tracks.
Inpainting workflowInput: Source video and mask. Notes: Two-stage; white mask regions are generated and black regions are preserved.
Outpainting workflowInput: Source video and target canvas. Notes: Two-stage; the padded region becomes the generated area.

LTX-2.3 specialized workflows

These workflows use LTX-2.3 assets. Do not substitute LTX-2.5 assets unless a workflow or model card explicitly establishes compatibility.

GoalOfficial starting pointRequired inputStatus
Video-to-Audio FoleyFoley V2A workflowMuted source video and a foley promptThe input video is preserved while the Foley track is generated.
Text-to-Audio FoleyLTX-2.3_T2A_Single_Stage_Distilled.json plus the Foley LoRAText promptDo not use the LTX-2.5 Text-to-Audio workflow as a substitute.
Dub-ItLTX-2.3_ICLoRA_DubIt_Two_Stage_Distilled.jsonSource video and replacement dialogueGenerates lip-synced replacement speech; it does not translate dialogue automatically.
RelightRelight guideExterior reference video, light-direction ball, and directional promptThe light-direction ball is part of the conditioning, not a decorative overlay.

Standalone LTX-2.5 LoRA workflows

Motion Transfer is a separate workflow and is not part of the VFX & Post-Production section.

CapabilitySupported surface and versionRequired workflow and inputLTX-2.3 versus LTX-2.5 boundary
Motion TransferLTX-2.5: ComfyUILTX-2.5_ICLoRA_Motion_Transfer_Distilled.json, look still, driving video, motion caption, and Motion Transfer IC-LoRA.Use the LTX-2.5 distilled model stack and the LTX-2.5 Motion Transfer IC-LoRA. LTX-2.3 assets are not supported, and the sparse-track Motion Control workflow is a different capability.

VFX & Post-Production workflows

The version boundary in each entry is mandatory. Do not substitute similarly named assets from another LTX release.

Surface: LTX-2.5 ComfyUI and native ltx-pipelines Python.

Workflow and input: In ComfyUI, use example_workflows/2.5/HDR_workflows/LTX-2.5_I2V_HDR_Two_Stage_Distilled.json with one EXR still. In Python, use an LTX-2.5 image-to-video pipeline with --image and matching --hdr.

Version boundary: Use the LTX-2.5 transformer, text encoder, video VAE, and spatial upscaler required by the selected surface. Do not substitute LTX-2.3 base-model assets. “Native HDR” means the documented EXR/HDR path, not general camera-RAW support.

Surface: LTX-2.5 ComfyUI.

Workflow and input: Use example_workflows/2.5/HDR_workflows/LTX-2.5_ICLoRA_Inpaint_HDR_Two_Stage_Distilled.json with an EXR sequence and mask.

Version boundary: This LTX-2.5 graph intentionally uses the specific LTX-2.3 In/Outpainting IC-LoRA. This is a workflow-defined exception only; keep the transformer, text encoder, VAEs, and upscaler on LTX-2.5 and do not mix other LTX-2.3 adapters or base assets.

Surface: LTX-2.3 ComfyUI and a version-matched legacy Python package.

Workflow and input: Use the published LTX-2.3 HDR workflow with an SDR source video, the LTX-2.3 HDR IC-LoRA, and the matching LTX-2.3 model stack. The current ltx_pipelines.hdr_ic_lora command is documented for LTX-2.5; use a version-matched older package to reproduce the legacy Python path.

Version boundary: Use only the LTX-2.3 LogC3 HDR adapter, scene embeddings, checkpoint, distilled LoRA/model, and spatial upsampler documented for this path. They are not substitutes for LTX-2.5 SDR-to-HDR assets.

Surface: LTX-2.5 ComfyUI and native Python.

Model card: LTX-2.5 22B IC-LoRA SDR to HDR.

Workflow and input: In ComfyUI, use LTX-2.5_ICLoRA_SDR_to_HDR_Distilled.json. In Python, run python -m ltx_pipelines.hdr_ic_lora. Both surfaces use an SDR source, ltx-2.5-22b-distilled-transformer-bf16.safetensors, a supported LTX-2.5 video VAE, ltx-2.5-22b-ic-lora-sdr-to-hdr-1.0.safetensors, and ltx-2.5-22b-ic-lora-sdr-to-hdr-scene-emb.safetensors. No prompt or runtime text encoder is required.

Version boundary: Use the LTX-2.5 distilled transformer and video VAE with the LTX-2.5 SDR-to-HDR adapter and embeddings. Both surfaces work internally in ACEScct and default to scene-linear ACEScg EXR output. Do not substitute the LTX-2.3 LogC3 adapter or scene embeddings.

Workflow-specific exception: LTX-2.5_V2V_TiledFusion_SDR_to_HDR.json instead loads the LTX-2.3 HDR IC-LoRA (ltx-2.3-22b-ic-lora-hdr-0.9) on the LTX-2.5 base. Follow each workflow’s model selection exactly.

Surface: LTX-2.5 ComfyUI. Do not infer LTX-2.3 or native Python compatibility.

Workflow and input: Use example_workflows/2.5/LTX-2.5_V2V_TiledFusion_Native_4K_8K.json (stage-1 viewport IC-LoRA and stage-2 refine IC-LoRA), or the Upscale, SDR→HDR, or HDR→HDR Tiled Fusion sibling. The graph supplies a version-matched model and IC-LoRA plus the guide node’s positive, negative, and latent inputs to LTXVTiledFusionSampler. Source audio is held fixed during generation and muxed into the saved video. LTXVTiledSampler is a separate implementation and is not a substitute.

Version boundary: Do not infer LTX-2.3 support from the node interface or universal LTX-2.5 compatibility from one example. Full HD, 4K, and 8K are selector values, not performance guarantees. The workflow filename’s Native label is unrelated to Native HDR.

Surface: LTX-2.5 ComfyUI and native Python. Do not infer LTX-2.3 compatibility.

Model card: LTX-2.5 22B IC-LoRA Alpha Gen.

Workflow and input: In ComfyUI, use LTX-2.5_V2V_ICLoRA_Single_Stage_Distilled.json with one RGB source video and ltx-2.5-22b-ic-lora-alpha-gen-0.9.safetensors at strength 1.0. Leave the prompt empty. In Python, run python -m ltx_pipelines.ic_lora with the same source and adapter, an empty prompt, and --skip-stage-2; pass width and height at twice the padded source dimensions so stage 1 returns the matte at native resolution.

Version boundary: The source is limited to 1920×1088, with dimensions divisible by 32 and frame count following 8n+1. Use 121 frames or fewer for the most predictable results; valid frame counts above 121 and up to 145 may work but can be less reliable, and longer clips are unsupported. The result is a frame-aligned grayscale matte. When used as an alpha channel, white keeps the model-selected area fully opaque, black makes the unselected area fully transparent, and gray produces partial opacity. The ComfyUI graph saves a standard video object and does not expose dedicated mask, RGBA, EXR, or alpha output sockets.

Surface: LTX-2.5 ComfyUI and native Python. Do not infer LTX-2.3 compatibility.

Model card: LTX-2.5 22B IC-LoRA Layout to Render.

Workflow and input: In ComfyUI, use LTX-2.5_ICLoRA_Layout_To_Render_Two_Stage_Distilled.json with a clay viewport or blocky playblast, an art-directed image made from its first frame, the LTX-2.5 distilled stack, and ltx-2.5-22b-ic-lora-layout-to-render-1.0.safetensors at strength 1.0. In Python, run python -m ltx_pipelines.ic_lora, pass the layout with --video-conditioning, pass the first-frame image with --image ... -1 ..., and enable --stage-2-ic-lora.

Version boundary: Width and height must be divisible by 64; 1920×1088 at 24 fps is a known-good recipe. Frame count snaps to 8k+1. The layout is the IC-LoRA guide and the image supplies art direction; optional mid/last keyframes can reinforce the look. This is not Motion Transfer. Neither surface preserves source audio: ComfyUI produces silent video, while Python generates new audio.

Surface: LTX-2.5 ComfyUI. Do not infer an LTX-2.3 or native Python path.

Model card: Refine Details. The separately published Restore adapter is not selected by this workflow.

Workflow and input: Use example_workflows/2.5/LTX-2.5_V2V_TiledFusion_Upscale.json with a source video and its audio, the Refine Details IC-LoRA, and the LTX-2.5 distilled stack. This is single-stage and does not use the spatial upscaler. The workflow provides Full HD, 4K, and 8K output selections.

Version boundary: This is a generative detail refiner, not pixel-accurate restoration; do not treat the LTX-2.3 or LTX-2.5 Pixel Spatial Upscaler as the same model. Test the selected output size on the target hardware.

Before you change a workflow

  1. Confirm the LTX version in the workflow, checkpoint, adapter card, and model files all match.
  2. Load the official workflow and run it without edits.
  3. Make one change at a time and test.
  4. Record the exact workflow source revision and the model-file identifiers used for a reproducible result.