DEV Community

AI-KSK
AI-KSK

Posted on Edited on Originally published at huggingface.co

Colourising Black-and-White Video: Chaining an Image Editor and a Video Model in One Graph

Colourising footage is a task where doing it frame by frame gives inconsistent colour — frame 40 comes out sepia, frame 41 comes out blue. Doing it at the video level gives coherent colour but muddy detail. This workflow does both, in one graph, in sequence.

The two-graph structure

The release ships a full automatic version (81 nodes) and a manual-prompt version (44 nodes).

The automatic version contains two model families in one graph:

Image stage — Qwen 2511
| Node | Value |
|---|---|
| UNETLoader | qwen_image_edit_2511_bf16.safetensors |
| CLIPLoader | qwen_2.5_vl_7b_fp8_scaled.safetensors, type qwen_image |
| VAELoader | qwen_image_vae.safetensors |
| EmptySD3LatentImage | 928 × 1664 |

Video stage — LTX 2.3
| Node | Value |
|---|---|
| CheckpointLoaderSimple | ltx-2.3-22b-dev.safetensors |
| LTXVAudioVAELoader | ltx-2.3-22b-dev.safetensors |
| LoraLoaderModelOnly | ltx-2.3-22b-distilled-lora-384.safetensors at 0.5 |
| EmptyLTXVLatentVideo | 960 × 544, 121 frames |
| KSamplerSelect | euler_ancestral_cfg_pp |

Why two stages

A single-frame image editor is good at deciding what colour things are. Give it a black-and-white frame and it will make a defensible choice: skin is skin-toned, sky is blue, foliage is green.

A video model is good at coherence over time. Give it 121 frames and it will keep the colour of an object consistent across them.

Neither alone gets both. The image stage establishes a colour reference; the video stage propagates it coherently.

The automatic version runs the image stage itself and feeds its output forward. The 44-node manual version expects you to supply the colour reference — which is the difference between 81 and 44 nodes.

The 960×544 / 121-frame geometry

This is the second workflow covered here that uses 121 frames at 960×544 rather than the more common 97 frames at 320×288.

For colourisation that geometry makes sense: colour decisions need resolution to be legible. At 320×288 a face is too few pixels for the model to commit to a skin tone confidently.

The dev checkpoint with 0.5 distillation, again

Same pattern as the transition workflow in an adjacent release: dev checkpoint plus distillation LoRA at 0.5, and the audio VAE matching the checkpoint exactly.

That configuration is emerging as this author's preferred baseline for tasks where the model needs to make its own decisions (transition paths, colour choices) rather than simply reproduce a reference.

Where this breaks

  1. Foreground/background separation. A black-and-white image has no colour cue separating a person from a wall behind them. The image stage has to guess, and it guesses wrong on lower-contrast footage.
  2. Historic footage. A 1950s street scene has colours the model has no strong prior for. You get plausible modern colour, not period-accurate colour.
  3. 121 frames is a hard segment limit. A longer clip needs multiple passes, and each pass makes its own colour decisions — so colour can shift at segment boundaries.
  4. Two model families means two sets of models to have on disk and two chances for a version mismatch.

Practical approach

Run the 44-node manual version first. Supplying the colour reference yourself on one frame tells you how much of the result is the model's judgement and how much is your input. Once that is understood, the automatic version becomes predictable.

Workflows

Two graphs, published on RunningHub with invite code rh-v1111:

  • LTX 2.3 IC LoRA video colourisation (automatic)
  • LTX 2.3 IC LoRA video colourisation (manual prompt)

Files

https://huggingface.co/AI-KSK/ksk-video-208-comfyui-workflows

Notes

  • Workflow files only. Model weights are not included; each keeps its own licence.
  • Tutorial exports, not a product.
  • Node counts and loader values quoted above are read from the published JSON files.
  • Independent AI-KSK release. Not affiliated with Lightricks, Alibaba or RunningHub.

Optional API platform - 88API (affiliate link)

Explore 88API for access to multiple AI model APIs. Available models, routes and prices are listed on the platform; Alipay top-ups are supported.

https://88api.ai/sign-up?aff=usx3

Disclosure: I may earn a commission from eligible purchases after you register through this referral link.

Top comments (1)

Collapse
 
alexshev profile image
Alex Shev •

Colourisation needs a temporal-consistency acceptance check in addition to attractive individual frames. A useful test set would include cuts, fast motion, skin tones, and repeated objects, then score both frame quality and identity or hue drift across adjacent frames. That makes it easier to spot a pipeline that is visually persuasive in stills but unstable in motion.