MULTY VIDO — Multi-Clip WAN 2.2 Movie Maker (up to 3 min continuous movies

# MULTY VIDO — Multi-Clip WAN 2.2 Movie Maker for ComfyUI

if this workflow saved you hours — a Buzz tip is appreciated

### Turn a list of prompts into ONE continuous movie — up to 3 MINUTES long — with consistent characters, locked colors, and fully automatic merging.

You write one prompt per scene, separated by ;. Each clip starts from the exact

last frame of the previous clip. Colors stay locked to clip 1 for the whole movie.

After every clip the workflow automatically updates a merged movie file — when the

last clip finishes, your full movie is simply waiting in output/finals/.

Movies up to ~3 minutes are absolutely possible (e.g. 36 clips × 5 seconds).

The only requirement: your prompts must be written as one continuous story —

every prompt must begin EXACTLY where the previous prompt ended (see section 4,

this is the heart of the method).

---

## 1. Requirements & Downloads

Hardware: a 24GB GPU runs everything below (fp8 models recommended for 24GB;

fp16 works with enough system RAM).

Diffusion models → ComfyUI/models/diffusion_models/

From [Comfy-Org/Wan_2.2_ComfyUI_Repackaged](https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged) split_files/diffusion_models/):

- wan2.2_i2v_high_noise_14B_fp16.safetensors (or fp8_scaled for 24GB GPUs)

- wan2.2_i2v_low_noise_14B_fp16.safetensors (or fp8_scaled)

- (if you use the fp8 files, just select them in the two Load Diffusion Model nodes)

Lightning 4-step LoRAs → ComfyUI/models/loras/

From [lightx2v/Wan2.2-Lightning](https://huggingface.co/lightx2v/Wan2.2-Lightning) (I2V A14B 4-steps rank64):

- high_noise LoRA → rename to Wan2.2-Lightning_I2V-A14B-4steps-lora_HIGH_fp16.safetensors

- low_noise LoRA → rename to Wan2.2-Lightning_I2V-A14B-4steps-lora_LOW_fp16.safetensors

Standard WAN files: umt5_xxl text encoder → models/text_encoders/,

wan_2.1_vae.safetensors → models/vae/

Recommended extensions (for the preview player): VideoHelperSuite.

## 2. Installation — step by step

1. Copy these THREE files into ComfyUI/custom_nodes/:

- multy_vido_v9_nodes.py

- prompt_splitter_by_semicolon.py

- drift_guard.py

2. Restart ComfyUI completely.

3. Verify the install: in the console you must see this line on startup:

[MULTY VIDO] custom nodes loaded, version 9.5.4

If you don't see it — the nodes file is not in the right folder.

4. Load MULTY_VIDO_PUBLIC.json into ComfyUI.

5. Load any start image into the Load Image node (this will be frame 1 of

your movie).

## 3. Running your first movie — step by step

1. Write your prompts in the Prompt Splitter node — one scene per clip,

each scene ending with ; (see section 4 for how to write them WELL).

2. Load a start image (Load Image node) — your movie starts from this frame.

The switch node "WITH / WITHOUT start image" should be on WITH image.

3. Set BatchCounter to 0. Do this before EVERY new movie. (The counter

auto-increments during the series; forgetting to reset it is the #1 user error.)

4. Set the run count — the number next to the Run button — to **exactly the

number of prompts you wrote** (7 prompts = run count 7).

5. Press Run and let it work. Each clip renders in a few minutes; you can

watch every finished clip in the preview player node.

6. Where your files appear:

- output/finals/multy_clip_MOVIE_LATEST.mp4 — the merged movie, updated

automatically after every clip. After the last clip this IS your full movie.

- output/finals/multy_clip_MOVIE_<timestamp>.mp4 — a stamped final copy,

created when the whole series completes.

- output/clips_<timestamp>/ — all the individual clips, archived next to

their movie. Nothing is ever deleted.

## 4. HOW TO WRITE PROMPTS — the rules that make or break the movie

This workflow chains clips: the last frame of clip N becomes the first frame of

clip N+1. The prompts must respect that chain. Three rules, in order of importance:

### RULE 1 — Every prompt must begin EXACTLY where the previous prompt ended

This is the single most important rule. The model can only continue what exists

in the frame — if prompt 3 ends with the couple walking toward the dance floor

and prompt 4 suddenly says "they are dancing", the model must invent the missing

connection, and that is where characters get swapped, hairstyles change, and

strange things happen.

Write the seams with explicit overlap:

- END each prompt by starting the next action ("...he gently squeezes her hand

and they both begin to rise from their chairs")

- BEGIN the next prompt from that exact state ("Continuing seamlessly, the couple

are standing up from their table...")

- Words like "Continuing seamlessly," / "Still in the same embrace," at the start

of a prompt work very well.

### RULE 2 — State the LIGHTING in every single prompt

Repeat the same lighting description in every prompt — ideally word for word

("warm golden sunset light", "soft candlelit restaurant lighting"). This

re-anchors brightness and atmosphere at the start of every clip and prevents the

slow drift toward washed-out or dark frames that ruins long chained videos.

The included DriftGuard node locks the color tone automatically — but the

lighting line in the prompt is what keeps the scene consistent. Use both.

### RULE 3 — Describe your main characters identically, word for word, every time

The samplers run on a FIXED seed (already configured — do not change it).

Fixed seed + identical character description = the same face, hair and clothing

in every clip. "the auburn-haired woman in the red evening dress" must appear

with those exact words in every prompt — every paraphrase ("the woman", "she")

re-rolls her appearance. Choose clothing that can't be misread (a white shirt +

black vest in a restaurant becomes a waiter; a navy suit stays a guest).

Bonus rule — one continuous action per clip. A clip is a few seconds.

Write ongoing states ("they are dancing"), not event sequences ("he offers his

hand, she accepts, they stand") — sequences don't complete in one clip.

The included example prompt (7-clip beach story) demonstrates all the rules —

read it before writing your own.

## 5. Making LONG movies (up to 3 minutes)

There is no hard limit on the number of clips — the movie length is simply

(number of prompts) × (clip length). Clip length is the length value in the

WanImageToVideo node: 81 ≈ 5s, 113 ≈ 7s (must be a multiple of 4, plus 1).

For a ~3 minute movie: 36 prompts × 5 seconds (or 26 × 7s). What matters:

- The prompts must be one continuous story following the rules of section 4

- Counter to 0, run count = number of prompts, press Run — and walk away;

a 3-minute movie is a few hours of rendering

- The rolling LATEST file lets you watch the movie grow while it renders

- For very long movies, keep the lighting line IDENTICAL throughout — over 30+

transitions, prompt discipline is what keeps the movie looking like ONE film

## 6. Using your own character LoRA (strongest identity)

For a persistent character across the whole movie, train a WAN 2.2 character

LoRA (two experts: high_noise + low_noise — e.g. with AI-Toolkit on the official

Wan-AI/Wan2.2-T2V-A14B-Diffusers base, one training run per expert).

Then: put both files in models/loras/, select them in the two

"OPTIONAL: your character LoRA" nodes, enable the nodes (select + Ctrl+B),

and put your trigger word in every prompt (Rule 3 still applies!).

## 7. Alternative: community merge checkpoints (e.g. Wan 2.2 "Remix")

The workflow also runs on community Wan 2.2 I2V merges:

1. Select the merge files in the two Load Diffusion Model nodes

2. Bypass the two ⚡ Lightning nodes (select + Ctrl+B) — merges usually have

the acceleration baked in; stacking it twice causes blur. (If a particular

merge does NOT include acceleration, keep Lightning active instead.)

3. Everything else stays the same

## 8. Troubleshooting

Video comes out BLURRY / smudged on the official models?

Usually a mismatch between your setup and the official checkpoints (Lightning

version, character LoRA trained on a different base, hardware quirks). Proven

fix: switch to a community merge like Wan 2.2 "Remix" — see section 7.

Merges are very forgiving and for many setups simply produce cleaner results.

Grainy / noisy output? The two ⚡ Lightning nodes are bypassed or not at

strength 1.0 — they must be ACTIVE at 1.0 on the official models (and BYPASSED

on merges).

Character melted / over-smoothed? Character LoRA strength too high or

trained on a different checkpoint family. Try 0.6–0.8, or retrain on the base

you generate with.

Colors or brightness drifting over a long movie? You skipped the lighting

line in some prompts (Rule 2) — it must appear in EVERY prompt. DriftGuard's

values (0.90/0.80/0.85) are tuned; change them only if you know why.

Characters change between clips? Rule 1 or Rule 3 was broken — check that

every prompt starts exactly where the previous ended, and that the character

descriptions are word-for-word identical. Also confirm the samplers' seed is

0/fixed (never randomize — the chaining depends on it).

The movie doesn't merge? Check the console for

[MULTY VIDO] custom nodes loaded, version 9.5.4 on startup. The LATEST movie

updates after every clip — if it doesn't, the nodes file isn't the one running.

"No numbered file found" error mid-series? BatchCounter wasn't reset to 0

before the series. Set it to 0 and start the series again.

---

Made with ❤️, a fixed seed, thousands of test clips — and one violinist who

stole the girl during testing.