<!-- Generated by tools/build_node_reference.py from the node definitions. Do not edit by hand. -->

# Video nodes

Text-to-video, image-to-video, temporal generation, and video HDR.

15 nodes. [All sections](README.md)

- [I2V Pipeline](#i2v-pipeline)
- [T2V Pipeline](#t2v-pipeline)
- [Video Assembler](#video-assembler)
- [Video Batch Decode](#video-batch-decode)
- [Video Cond Merge](#video-cond-merge)
- [Video Export](#video-export)
- [Video Frame Router](#video-frame-router)
- [Video HDR Conditioner](#video-hdr-conditioner)
- [Video HDR Decode](#video-hdr-decode)
- [Video Latent Noise](#video-latent-noise)
- [Video Loader](#video-loader)
- [Video Mask Propagator](#video-mask-propagator)
- [Video Model Info](#video-model-info)
- [Video Prompt Builder](#video-prompt-builder)
- [Video Sampler](#video-sampler)

## I2V Pipeline

`RadianceI2VPipeline`

End-to-end image-to-video generation pipeline with motion control.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `model` | MODEL |  |  | Video diffusion model. Image models are rejected; a model with an image-concat input (Wan I2V) enables concat_channels. |
| `clip` | CLIP |  |  | Text encoder matching the model, used for both prompts. |
| `vae` | VAE |  |  | VAE matching the model. Encodes the reference image, sets the latent shape and decodes preview_frames. |
| `reference_image` | IMAGE |  |  | Start frame (display-referred sRGB, one image). Its width and height set the video size, rounded down to the VAE's spatial compression. |
| `positive_prompt` | string | `smooth camera motion, cinematic HDR, 4K` | multi-line text | What should happen in the shot. ', <n> nits HDR, <gamut>' is appended when peak_nits is above 100. |
| `negative_prompt` | string | `watermark, blurry, flickering, sdr` | multi-line text | What to steer away from, encoded with the same text encoder. |
| `frames` | int | 25 | 1 to 512 | Requested frames. The latent holds ceil(frames / temporal compression) frames; use a multiple of the compression plus 1 (e.g. 25, 49, 81) to get exactly this count back. |
| `seed` | int | 0 | 0 to 2147483648 | Seed for the initial noise and the sampler. |
| `dit_config` (optional) | string | `{}` |  | JSON from RadianceVideoModelInfo. When it carries a model_name, that model's defaults replace steps, cfg, sampler_name and scheduler. |
| `character_conditioning` (optional) | CONDITIONING |  |  | Optional conditioning whose tokens are appended to the positive prompt at weight 0.75. Skipped if its embedding width differs. |
| `cfg_schedule_json` (optional) | string |  |  | JSON float array. Only the first value is used, as a static CFG override; CFG does not vary per step. |
| `i2v_strategy` (optional) | choice | `auto` | `auto`, `first_frame_lock`, `concat_channels`, `clip_vision_inject`, `prepend_latent` | auto: concat_channels when the model has an image-concat input (Wan 2.1 I2V, Wan 2.2 I2V 14B), else first_frame_lock. clip_vision_inject needs clip_vision_output connected. |
| `clip_vision_output` (optional) | CLIP_VISION_OUTPUT |  |  | From CLIP Vision Encode. Added to both conditionings as clip_vision_output, which Wan 2.1 I2V reads (clip_fea). |
| `image_strength` (optional) | float | 0.85 | 0 to 1, step 0.01 | first_frame_lock and prepend_latent only: blend weight of the image latent into the first latent frame; also lowers denoise to 1 - 0.4 x strength. Ignored by concat_channels and clip_vision_inject. |
| `motion_strength` (optional) | float | 0.5 | 0 to 1, step 0.01 | first_frame_lock and prepend_latent only: scales the starting noise to 0.2 + 0.8 x strength of its unit amplitude. Ignored by concat_channels and clip_vision_inject. |
| `steps` (optional) | int | 0 | 0 to 200 | Sampling steps. 0 uses the model default (the LTX-Video preset's 25 when no dit_config is connected). Ignored when dit_config carries a model_name. |
| `cfg` (optional) | float | 0 | 0 to 30, step 0.1 | Guidance scale. 0 uses the model default (the LTX-Video preset's 3.5 when no dit_config is connected). cfg_schedule_json overrides it. |
| `sampler_name` (optional) | choice | `euler` | ComfyUI's samplers | ComfyUI sampler. Ignored when dit_config carries a model_name. |
| `scheduler` (optional) | choice | `normal` | ComfyUI's schedulers | ComfyUI sigma scheduler. Ignored when dit_config carries a model_name. |
| `peak_nits` (optional) | choice | `1000` | `100`, `203`, `400`, `600`, `1000`, `4000`, `10000` | Adds the selected peak brightness to the prompt; 100 requests an SDR look. No pixel change. |
| `target_gamut` (optional) | choice | `BT.2020` | `BT.2020`, `P3-D65`, `P3-DCI`, `BT.709`, `ACEScg` | Adds the selected gamut descriptor to the prompt at every peak brightness. No pixel conversion. |
| `hdr_eotf` (optional) | choice | `PQ (ST.2084)` | `PQ (ST.2084)`, `HLG (BT.2100)`, `Linear`, `sRGB / BT.1886` | Adds the selected transfer-function descriptor to the prompt. No pixel encoding. |

**Outputs**

| Output | Type |
| :--- | :--- |
| `video_latent` | LATENT |
| `preview_frames` | IMAGE |
| `pipeline_report` | STRING |

## T2V Pipeline

`RadianceT2VPipeline`

End-to-end text-to-video generation pipeline with HDR support.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `model` | MODEL |  |  | Video diffusion model. Image models (2-D latents) are rejected. |
| `clip` | CLIP |  |  | Text encoder matching the model, used for both prompts. |
| `vae` | VAE |  |  | VAE matching the model. Its compression sets the latent shape and it decodes preview_frames. |
| `positive_prompt` | string | `cinematic HDR video, stunning visuals, 4K, film grain` | multi-line text | What to generate. HDR descriptors from peak_nits, target_gamut and hdr_eotf are appended unless hdr_strength is 0. |
| `negative_prompt` | string | `watermark, blurry, low quality, sdr, flickering` | multi-line text | What to steer away from, encoded with the same text encoder. |
| `width` | int | 768 | 64 to 4096, step 8 | Output width in pixels, rounded down to a multiple of the VAE's spatial compression. |
| `height` | int | 512 | 64 to 4096, step 8 | Output height in pixels, rounded down to a multiple of the VAE's spatial compression. |
| `frames` | int | 25 | 1 to 512 | Requested frames. The latent holds ceil(frames / temporal compression) frames; use a multiple of the compression plus 1 (e.g. 25, 49, 81) to get exactly this count back. |
| `seed` | int | 0 | 0 to 2147483648 | Seed for the initial noise and the sampler. |
| `dit_config` (optional) | string | `{}` |  | JSON from RadianceVideoModelInfo. When it carries a model_name, that model's defaults replace steps, cfg, sampler_name and scheduler. |
| `character_conditioning` (optional) | CONDITIONING |  |  | Optional conditioning whose tokens are appended to the positive prompt at weight 0.75. Skipped if its embedding width differs. |
| `cfg_schedule_json` (optional) | string |  |  | JSON float array from RadianceAudioCFGSchedule. Only the first value is used, as a static CFG override; CFG does not vary per step. |
| `steps` (optional) | int | 0 | 0 to 200 | Sampling steps. 0 uses the model default (the LTX-Video preset's 25 when no dit_config is connected). Ignored when dit_config carries a model_name. |
| `cfg` (optional) | float | 0 | 0 to 30, step 0.1 | Guidance scale. 0 uses the model default (the LTX-Video preset's 3.5 when no dit_config is connected). cfg_schedule_json overrides it. |
| `sampler_name` (optional) | choice | `euler` | ComfyUI's samplers | ComfyUI sampler. Ignored when dit_config carries a model_name. |
| `scheduler` (optional) | choice | `normal` | ComfyUI's schedulers | ComfyUI sigma scheduler. Ignored when dit_config carries a model_name. |
| `peak_nits` (optional) | choice | `1000` | `100`, `203`, `400`, `600`, `1000`, `4000`, `10000` | Adds '<n> nits HDR' to the prompt. Prompt text only, no pixel change; 100 requests an SDR look. |
| `target_gamut` (optional) | choice | `BT.2020` | `BT.2020`, `P3-D65`, `P3-DCI`, `BT.709`, `ACEScg` | Adds the selected gamut descriptor to the prompt. Prompt text only, no pixel conversion. |
| `hdr_eotf` (optional) | choice | `PQ (ST.2084)` | `PQ (ST.2084)`, `HLG (BT.2100)`, `Linear`, `sRGB / BT.1886` | Adds a transfer-function descriptor to the prompt. Prompt text only, no pixel encoding. |
| `hdr_strength` (optional) | float | 0.5 | 0 to 1, step 0.05 | Prompt weight of the appended HDR descriptors (gamut, EOTF, peak nits), as (text:weight) with weight = 2 x strength: 0.5 is neutral, 1.0 doubles their emphasis, 0 leaves them out. |

**Outputs**

| Output | Type |
| :--- | :--- |
| `video_latent` | LATENT |
| `preview_frames` | IMAGE |
| `positive_cond` | CONDITIONING |
| `pipeline_report` | STRING |

## Video Assembler

`RadianceVideoAssembler`

Assemble processed video frames back into a temporal sequence.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `frame` | IMAGE |  |  | Frame (or batch) to append to this session's buffer. Buffers live in memory and are lost on restart. |
| `session_key` | string | `video_session_0` |  | Name of the accumulation buffer. Use a different key per clip being assembled in parallel. |
| `expected_total_frames` | int | 24 |  | Output is complete, and the buffer cleared, once this many inputs have arrived. Counts executions, not images, if frame is a batch. |
| `flush` (optional) | boolean | off |  | Force output of accumulated frames now, even if incomplete |
| `reset` (optional) | boolean | off |  | Clear accumulated frames for this session key |

**Outputs**

| Output | Type |
| :--- | :--- |
| `video_image` | IMAGE |
| `frames_accumulated` | INT |
| `is_complete` | BOOLEAN |

## Video Batch Decode

`RadianceVideoBatchDecode`

Batch-decode video latents to pixel frames with memory management.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `vae` | VAE |  |  | VAE matching the model that produced the latent. |
| `latent` | LATENT |  |  | Video latent from a sampler or pipeline. A 5-D latent is decoded through the VAE's temporal path. |
| `dit_config` (optional) | string | `{}` |  | JSON from RadianceVideoModelInfo. Its compression values are used only if the VAE does not report its own. Its latent_scale is not applied: a ComfyUI sampler already returns the latent in VAE space. |
| `tile_decode` (optional) | boolean | off |  | Route the decode through the VAE's own tiled entry point (comfy.sd.VAE.decode_tiled) to cut peak VRAM on large videos. Leave off: an untiled decode already falls back to tiling by itself when it runs out of memory. |
| `tile_overlap` (optional) | int | 64 | 0 to 256 | Pixel overlap between spatial tiles (higher = smoother seams). Only read when tile_decode is on. |
| `output_linear` (optional) | boolean | off |  | Convert the VAE's display-referred sRGB frames to scene-linear (inverse sRGB transfer). Off: sRGB clamped to [0, 1], as VAE Decode. |

**Outputs**

| Output | Type |
| :--- | :--- |
| `frames` | IMAGE |
| `frame_count` | INT |
| `decode_report` | STRING |

## Video Cond Merge

`RadianceVideoCondMerge`

Merge multiple video conditioning signals into a unified tensor.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `text_conditioning` | CONDITIONING |  |  | Base conditioning, usually the encoded prompt. The output keeps its entries; the other inputs are merged into them. |
| `merge_mode` | choice | `concat` | `concat`, `weighted`, `priority` | concat: append the other inputs' tokens (times their weights) after the text tokens. weighted: weighted average of the token tensors. priority: keep the text tokens, only copy missing dict keys from the others. |
| `character_conditioning` (optional) | CONDITIONING |  |  | Optional conditioning (e.g. a character or identity embedding) merged per merge_mode. |
| `hdr_conditioning` (optional) | CONDITIONING |  |  | Optional conditioning (e.g. from RadianceVideoHDRConditioner) merged per merge_mode. |
| `text_weight` (optional) | float | 1 | 0 to 2, step 0.05 | Weight of text_conditioning in weighted mode. concat and priority keep the text tokens unscaled. |
| `character_weight` (optional) | float | 0.75 | 0 to 2, step 0.05 | Multiplier on the character tokens in concat and weighted modes. Ignored in priority mode. |
| `hdr_weight` (optional) | float | 0.5 | 0 to 2, step 0.05 | Multiplier on the HDR tokens in concat and weighted modes. Ignored in priority mode. |

**Outputs**

| Output | Type |
| :--- | :--- |
| `merged_conditioning` | CONDITIONING |
| `merge_report` | STRING |

## Video Export

`RadianceVideoExport`

Pass decoded video frames through, HDR-encode them, or write them as a 32-bit float EXR sequence or a small GIF preview. No movie codecs: use a video writer node.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `frames` | IMAGE |  |  | Decoded frames. EXR writes the values unchanged as 32-bit float; the GIF clamps to [0, 1] display-referred. |
| `mode` | choice | `passthrough` | `passthrough`, `hdr_decode`, `exr_sequence`, `preview_gif` | passthrough: return frames. hdr_decode: run RadianceVideoHDRDecode (Reinhard, PQ out) and return the HDR signal. exr_sequence / preview_gif: also write files. Only the last two write to disk. |
| `hdr_metadata_json` (optional) | string | `{"peak_nits":1000,"eotf":"PQ (ST.2084)"}` |  | hdr_decode only. Its peak_nits and gamut (BT.2020 if absent) are used; the eotf key is ignored, output is always PQ. |
| `output_folder` (optional) | string |  |  | Empty: ComfyUI's output folder. |
| `filename_prefix` (optional) | string | `radiance_video` |  | File name start: <prefix>_<6-digit frame>.exr, or <prefix>_preview.gif (overwritten on each run). |
| `fps` (optional) | float | 24 | 1 to 120 | GIF playback rate (frame duration 1000 / fps ms). preview_gif only. |
| `frame_offset` (optional) | int | 0 |  | Added to the frame number in EXR file names. exr_sequence only. |

**Outputs**

| Output | Type |
| :--- | :--- |
| `frames` | IMAGE |
| `frame_count` | INT |
| `export_report` | STRING |

## Video Frame Router

`RadianceVideoFrameRouter`

Route individual video frames to different processing branches.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `video_image` | IMAGE |  |  | Frame batch to pick from. Also returned unchanged on passthrough. |
| `frame_index` | int | 0 | 0 to 4096 | 0-based frame to extract. Past the end it wraps or clamps to the last frame, per wrap_index. |
| `wrap_index` (optional) | boolean | on |  | If frame_index >= total_frames, wrap around (modulo). Off: clamp to the last frame. |

**Outputs**

| Output | Type |
| :--- | :--- |
| `frame_image` | IMAGE |
| `frame_index` | INT |
| `total_frames` | INT |
| `passthrough` | IMAGE |

## Video HDR Conditioner

`RadianceVideoHDRConditioner`

Condition a video model on HDR metadata for luminance-aware sampling.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `positive` | CONDITIONING |  |  | Encoded positive prompt. The HDR descriptor embeddings are concatenated onto each entry (needs clip). |
| `peak_nits` | choice | `1000` | `100`, `203`, `400`, `600`, `1000`, `4000`, `10000` | Mastering peak in nits: adds a luminance descriptor to the tokens and is stored as peak_nits in hdr_metadata_json for RadianceVideoHDRDecode. |
| `target_gamut` | choice | `BT.2020` | `BT.2020`, `P3-D65`, `P3-DCI`, `BT.709`, `ACEScg`, `ACES2065-1` | Adds a gamut descriptor to the tokens and is stored as gamut in hdr_metadata_json (RadianceVideoHDRDecode converts to it). |
| `eotf` | choice | `PQ (ST.2084)` | `PQ (ST.2084)`, `HLG (BT.2100)`, `Linear`, `sRGB / BT.1886` | Adds a transfer-function descriptor to the tokens and is stored in hdr_metadata_json. RadianceVideoHDRDecode uses its own output_eotf. |
| `clip` (optional) | CLIP |  |  | Text encoder used for positive. Required for the descriptors to reach the model. |
| `camera_move` (optional) | choice | `None` | `Handheld documentary`, `Locked off cinematic`, `Slow push-in`, `Drone aerial`, `Tracking shot`, `Static time-lapse`, `None` | Adds camera-movement words to the descriptor tokens. None adds nothing. |
| `mood` (optional) | choice | `None` | `Golden hour`, `Blue hour / dusk`, `Night`, `Overcast flat`, `High contrast`, `Neon / cyberpunk`, `Natural daylight`, `None` | Adds lighting-mood words to the descriptor tokens. None adds nothing. |
| `extra_hdr_prompt` (optional) | string |  | multi-line text | Additional HDR descriptors appended to conditioning tokens |
| `inject_metadata_embedding` (optional) | boolean | on |  | Store the HDR metadata in the conditioning for Radiance nodes. No ComfyUI model reads it. |
| `token_strength` (optional) | float | 1 | 0 to 2, step 0.05 | Scale of the concatenated descriptor embeddings (1.0 = as encoded). Needs clip. |

**Outputs**

| Output | Type |
| :--- | :--- |
| `positive` | CONDITIONING |
| `hdr_metadata_json` | STRING |

## Video HDR Decode

`RadianceVideoHDRDecode`

Encode decoded sRGB video frames (IMAGE, not latents) to an HDR signal at the metadata's peak nits and gamut, with PQ or HLG output and an SDR preview.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `image` | IMAGE |  |  | Decoded video frames, display-referred sRGB Rec.709 in [0, 1]. Linearised with a pure 2.2 gamma; 1.0 is mapped to peak_nits. |
| `hdr_metadata_json` | string | `{"peak_nits":1000,"gamut":"BT.2020","eotf":"PQ (ST.2084)"}` |  | JSON from RadianceVideoHDRConditioner or manually entered |
| `tonemap` | choice | `Reinhard` | `Reinhard`, `Linear clip`, `Pass-through` | Reinhard: extended Reinhard whose white point is the brightest input (1.0 lifted by a positive exposure_compensation_ev), so that value lands exactly on peak_nits and the highlights above it roll off; at 0 EV or less there is nothing to compress. Linear clip: clamp at 10,000 nits. Pass-through: no curve (clamped at 10,000 nits by the encode). |
| `exposure_compensation_ev` (optional) | float | 0 | -6 to 6, step 0.1 | EV adjustment before tone-mapping |
| `output_eotf` (optional) | choice | `PQ (ST.2084)` | `PQ (ST.2084)`, `HLG (BT.2100)`, `Linear`, `sRGB / BT.1886` | Encoding of hdr_image. PQ: ST 2084 code values (1.0 = 10,000 nits). HLG: BT.2100 OETF. Linear: clamped linear light normalised to 10,000 nits. sRGB / BT.1886: the sRGB curve on light relative to peak_nits (1.0 = peak), an SDR signal. |
| `sdr_preview_nits` (optional) | float | 100 | 1 to 203 | Nits shown as white-ish mid-range in sdr_preview: light is measured in units of this value and rolled off with extended Reinhard so the brightest input reaches display white (1.0). Lower = brighter preview. |
| `gamut_clip` (optional) | boolean | on |  | Clamp to [0, 1] of the encode container after the primaries conversion (negatives from out-of-gamut colours, values above peak) |

**Outputs**

| Output | Type |
| :--- | :--- |
| `hdr_image` | IMAGE |
| `sdr_preview` | IMAGE |
| `decode_report` | STRING |

## Video Latent Noise

`RadianceVideoLatentNoise`

Correctly shaped, seeded i.i.d. Gaussian latent noise for a video model (independent per frame, as every video sampler expects).

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `dit_config` | string | `{}` |  | JSON from RadianceVideoModelInfo |
| `width` | int | 512 | 64 to 4096, step 8 | Target frame width in pixels. Divided (rounded down) by the spec's spatial compression to size the latent. |
| `height` | int | 512 | 64 to 4096, step 8 | Target frame height in pixels. Divided (rounded down) by the spec's spatial compression to size the latent. |
| `frames` | int | 25 | 1 to 512 | Target pixel frames. The latent gets ceil(frames / temporal compression) frames; use a multiple of the compression plus 1 (e.g. 25, 49) to decode back to exactly this count. |
| `batch_size` | int | 1 | 1 to 16 | Number of independent noise samples in the batch. |
| `seed` | int | 0 | 0 to 2147483648 | Seed for the CPU noise generator; the same seed and shape give identical noise. |
| `noise_scale` (optional) | float | 1 | 0.01 to 4, step 0.01 | Multiply noise standard deviation (1.0 = unit Gaussian) |

**Outputs**

| Output | Type |
| :--- | :--- |
| `noise_latent` | LATENT |
| `shape_report` | STRING |

## Video Loader

`RadianceVideoLoader`

Video loader v3.3 — for LTX 2.3, Wan, HunyuanVideo, etc. Supports Baked/standalone VAE, optional Audio VAE, and optional latent upscale model, on top of the universal loader features.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `preset` | choice | `Custom` | `Custom`, `CogVideoX`, `Cosmos World`, `HunyuanVideo`, `LTX Video`, `LTX Video (Low VRAM)`, `LTX Video 2.3`, `LTX Video 2.3 (Low VRAM)`, `LTX Video 2.5`, `LTX Video 2.5 (Low VRAM)`, and 8 more | Quick-configure for common architectures. Overrides model_type, dtypes, offload_mode, and hints which CLIP slots are needed. |
| `unet_name` | choice |  | files found in the matching models or input folder | Main diffusion model (UNET / DiT / Transformer). For WAN 2.2, select either the high_noise or low_noise file — the companion expert is detected automatically. The 'model' output always carries the high_noise expert and 'model_low_noise' always carries the low_noise expert, regardless of which file is selected. |
| `weight_dtype` | choice | `default` | `default`, `fp8_e4m3fn`, `fp8_e5m2`, `fp16`, `bf16`, `fp32` | UNET weight precision. fp8_e4m3fn saves ~40% VRAM vs fp16. |
| `model_type` | choice | `Auto-Detect` | `Auto-Detect`, `hunyuan_video`, `wan`, `ltxv`, `ltxav`, `wan_ti2v`, `cosmos`, `cogvideox`, `mochi`, `minimax`, and 2 more | 'Auto-Detect' reads the checkpoint's key names to determine architecture. Override manually if detection fails. |
| `vae_name` | choice | `Baked VAE (from UNET)` | `Baked VAE (from UNET)` | VAE for encoding/decoding latents. 'Baked VAE (from UNET)' extracts it from the checkpoint. |
| `audio_vae_name` (optional) | choice | `None` | `None`, `Baked Audio VAE (from UNET)` | Audio VAE for LTX 2.3. Choose 'Baked' or a standalone safetensors file. |
| `upscale_model_name` (optional) | choice | `None` | `None` | Latent Upscale Model (e.g. for LTX 2.3 or HunyuanVideo). |
| `clip_l` (optional) | choice | `None` | `None` | CLIP-L (text encoder). Used by: SD1.5, SDXL, Flux, SD3. |
| `clip_g` (optional) | choice | `None` | `None` | CLIP-G (text encoder). Used by: SDXL, SD3, SD3.5. |
| `t5xxl` (optional) | choice | `None` | `None`, `Baked (from UNET)` | T5-XXL (text encoder). Used by: Flux, SD3, SD3.5, Wan, PixArt, LTX (pre-2.3). 'Baked (from UNET)' loads it from the main checkpoint -- AuraFlow ships no standalone text encoder file. |
| `llm_encoder` (optional) | choice | `None` | `None` | LLM encoder. Used by: HunyuanVideo (Llava-Llama3), LTX 2.3 (Gemma 3), Lumina2 (Gemma-2), Z-Image (Qwen3), Flux.2 (Mistral-3/Qwen3). |
| `text_projection` (optional) | choice | `None` | `None`, `Baked (from UNET)` | Text projection matrix. Used by: LTX 2.3 (with Gemma 3 llm_encoder). 'Baked (from UNET)' loads it from the main LTX 2.3 checkpoint, like the native LTXV Audio Text Encoder Loader. |
| `clip_dtype` (optional) | choice | `default` | `default`, `fp16`, `bf16`, `fp8_e4m3fn`, `fp32` | CLIP weight precision. Independent from UNET. For Flux T5XXL: fp8 saves ~4.7 GB vs fp16. |
| `offload_mode` (optional) | choice | `none` | `none`, `cpu_offload`, `sequential` | none = GPU only. cpu_offload = CLIP loaded to CPU RAM. sequential = enable ComfyUI sequential CPU offload (8–12 GB GPUs). |
| `lora_stack` (optional) | LORA_STACK |  |  | Accept a LORA_STACK from RadianceLoraStack node. |
| `check_vram` (optional) | choice | `On` | `On`, `Off` | Estimate VRAM before load and warn if tight. |
| `use_cache` (optional) | choice | `On` | `On`, `Off` | Cache loaded models. Skips disk I/O when re-running with the same files. Cache auto-invalidates if files change. |
| `lora_on_error` (optional) | choice | `raise` | `warn`, `raise` | 'warn' skips failed LoRA and continues. 'raise' stops execution. |
| `auto_download` (optional) | boolean | on |  | If a selected model is missing and is one Radiance knows, download it on first run from its pinned Hugging Face source, checked against its SHA-256 before it is installed (large: 4 to 60 GB). Gated repositories (FLUX.2-dev, FLUX.2-klein 9B, LTX-2.5) need their licence accepted on Hugging Face and HF_TOKEN set. RADIANCE_ALLOW_DOWNLOADS=0 always stops downloads. |

**Outputs**

| Output | Type |
| :--- | :--- |
| `model` | MODEL |
| `model_low_noise` | MODEL |
| `clip` | CLIP |
| `vae` | VAE |
| `audio_vae` | VAE |
| `lora_stack` | LORA_STACK |
| `upscale_model` | LATENT_UPSCALE_MODEL |
| `model_meta` | STRING |

## Video Mask Propagator

`RadianceVideoMaskPropagator`

Fill empty frames of a mask sequence by warping neighbouring keyframe masks along optical flow. Frames that already contain a mask are kept as keyframes.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `masks` | MASK |  |  | Mask sequence, one per frame. Frames with any mask content are keyframes; empty frames are filled by propagation. |
| `flow_vectors` | IMAGE |  |  | 32-bit flow vectors from Radiance Optical Flow. |
| `propagation_mode` | choice | `Bidirectional` | `Forward`, `Backward`, `Bidirectional` | Forward carries masks from earlier frames, Backward from later frames. Bidirectional runs both and keeps the union (maximum) on filled frames. |

**Outputs**

| Output | Type |
| :--- | :--- |
| `propagated_masks` | MASK |

## Video Model Info

`RadianceVideoModelInfo`

Display configuration and parameter info for a loaded video model.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `model` | MODEL |  |  | Video diffusion model to inspect. It is passed through unchanged on the model output. |
| `model_preset` | choice | `LTX-Video (128ch)` | `SD-VAE (4ch)`, `SDXL-VAE (4ch)`, `LTX-Video (128ch)`, `HunyuanVideo (16ch)`, `Wan2.1 (16ch)`, `Wan2.2-T2V-14B (16ch)`, `Wan2.2-I2V-14B (16ch)`, `Wan2.2-TI2V-5B (48ch)`, `HunyuanVideo-1.5 (32ch)`, `CogVideoX (16ch)`, and 1 more | Latent spec used when auto-detection finds nothing. Detection reads the model class name (ltx, hunyuan, wan, cogvideo, mochi) and the latent channel count the model reports. A preset from the detected family that matches the channels is kept; otherwise the matching one is used. |
| `override_channels` (optional) | int | 0 | 0 to 512 | Replace the latent channel count written to dit_config. 0 keeps the preset's value. |
| `override_latent_scale` (optional) | float | 0 | 0 to 10 | Replace the preset's latent_scale in dit_config. Reported for reference; Video Batch Decode does not apply it, since sampler output is already in VAE space. 0 keeps the preset's value. |
| `print_info` (optional) | boolean | off |  | Also write the info report to the ComfyUI console log. |

**Outputs**

| Output | Type |
| :--- | :--- |
| `model` | MODEL |
| `dit_config` | STRING |
| `info_report` | STRING |

## Video Prompt Builder

`RadianceVideoPromptBuilder`

Build structured video prompts combining text, style, and motion cues.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `subject` | string | `a person walking through a neon-lit cityscape` |  | Main subject and action. Placed first in the positive prompt. |
| `peak_nits` | choice | `1000` | `100`, `203`, `400`, `600`, `1000`, `4000`, `10000` | Adds a peak-luminance descriptor (e.g. '1000 nits HDR10 ...') to the prompt. Prompt text only. |
| `target_gamut` | choice | `BT.2020` | `BT.2020`, `P3-D65`, `P3-DCI`, `BT.709`, `ACEScg`, `ACES2065-1` | Adds a gamut descriptor to the prompt. Prompt text only. |
| `eotf` | choice | `PQ (ST.2084)` | `PQ (ST.2084)`, `HLG (BT.2100)`, `Linear`, `sRGB / BT.1886` | Adds a transfer-function descriptor to the prompt. Prompt text only. |
| `camera_move` (optional) | choice | `Slow push-in` | `Handheld documentary`, `Locked off cinematic`, `Slow push-in`, `Drone aerial`, `Tracking shot`, `Static time-lapse`, `None` | Adds camera-movement words after the mood words. None adds nothing. |
| `mood` (optional) | choice | `Neon / cyberpunk` | `Golden hour`, `Blue hour / dusk`, `Night`, `Overcast flat`, `High contrast`, `Neon / cyberpunk`, `Natural daylight`, `None` | Adds lighting-mood words right after the subject. None adds nothing. |
| `style_suffix` (optional) | string | `photorealistic, 8K, film grain, anamorphic lens` | multi-line text | Free text appended at the end of the positive prompt. |
| `suppress_artefacts` (optional) | boolean | on |  | On: negative_prompt is a fixed list of common video artefacts. Off: negative_prompt is empty. |
| `print_prompt` (optional) | boolean | off |  | Also write both prompts to the ComfyUI console log. |

**Outputs**

| Output | Type |
| :--- | :--- |
| `positive_prompt` | STRING |
| `negative_prompt` | STRING |

## Video Sampler

`RadianceVideoSampler`

Run the diffusion sampler to generate video latents from pre-built noise and conditioning.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `model` | MODEL |  |  | Video diffusion model to sample with. |
| `positive` | CONDITIONING |  |  | Positive (prompt) conditioning. |
| `negative` | CONDITIONING |  |  | Negative conditioning, used by classifier-free guidance. |
| `latent_noise` | LATENT |  |  | Noise latent, e.g. from RadianceVideoLatentNoise. It is used as both the noise and the start latent, so its shape must match the model. |
| `steps` | int | 25 | 1 to 200 | Sampling steps. Ignored when dit_config carries a model_name. |
| `cfg` | float | 7 | 0 to 30, step 0.1 | Classifier-free guidance scale. Replaced by the model default when dit_config carries a model_name, then by the first value of cfg_schedule_json. |
| `sampler_name` | choice | `euler` | ComfyUI's samplers | ComfyUI sampler. Ignored when dit_config carries a model_name. |
| `scheduler` | choice | `normal` | ComfyUI's schedulers | ComfyUI sigma scheduler. Ignored when dit_config carries a model_name. |
| `seed` | int | 0 | 0 to 2147483648 | Seed for the sampler's own noise (ancestral and SDE samplers). The initial noise comes from latent_noise. |
| `dit_config` (optional) | string | `{}` |  | JSON from RadianceVideoModelInfo — when connected, overrides steps/cfg/sampler/scheduler with model-specific defaults. |
| `cfg_schedule_json` (optional) | string |  |  | JSON float array from RadianceAudioCFGSchedule — first value overrides CFG |
| `denoise` (optional) | float | 1 | 0 to 1, step 0.01 | Fraction of the noise schedule to run (1.0 = full). The start latent is latent_noise itself, so below 1.0 this is not a video-to-video strength. |

**Outputs**

| Output | Type |
| :--- | :--- |
| `samples` | LATENT |
| `sampler_report` | STRING |