← sources

Try this workflow, please. It cuts degradation in half just by adding it.

source: reddit (reddit:StableDiffusion)published 2026-10-29 01:55ingested 2026-10-30 02:38

retained text

I'm still sick.

slowing degradation by about 60% while having virtually no impact on motion.

[https://www.reddit.com/r/StableDiffusion/comments/1xi7cng/comment/19ruh5kx0vs90/?utm\_source=share&utm\_medium=web3x&utm\_name=web3xcss&utm\_term=1&utm\_content=share\_button](https://www.reddit.com/r/StableDiffusion/comments/1xi7cng/comment/19ruh5kx0vs90/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button)

Please try it. Test it. Fuck with the settings. Leave comments on the github or in this post. Be my little beta testers.

[https://github.com/roycho87/clown\_context\_suite/blob/main/example\_workflows/mad\_clown\_inventor\_example.json](https://github.com/roycho87/clown_context_suite/blob/main/example_workflows/mad_clown_inventor_example.json)

it should work pretty much like regular motion context. the only difference will be how you join the videos and when you change your scene you need to disable it while the scene changes and change the source latent to the beginning of that new scene in the clown keyframe node.

>eli5 explaination.

>when you generate something you're basically stacking all these dominoes in a line, and when you generate it, all those dominoes lean on eachother to form a picture.

>Motion context, as it has been used, has been taking the end of that domino chain and restarting a new chain, but we didn't realize that everything that made that picture good was in that very first domino.

>so i changed it so that when we take the end of the domino chain to form a new picture, i also pick up that first domino and bring it with so that the picture still comes out good.

Info from the readme.

[https://github.com/roycho87/clown\_context\_suite](https://github.com/roycho87/clown_context_suite)

# Clown Keystone and Clown Late Lock

Every shot in a chain is made from the one before it, so small colour and sharpness changes build up. Clown Keystone puts a latent from clip 1 into every shot, so each one is pulled back to how clip 1 looked.

H3 latents come in 17-frame blocks of 1 + 4 + 4 + 4 + 4 frames. The first latent of each block is a single frame on its own:

* **keystone** = clip 1's first latent (frame 0).
* **capstone** = clip 1's last single-frame latent (frame 187 of a 192-frame clip).

Keystone only swaps those single-frame latents, so nothing else in the shot moves.

|Setting|Meaning|
|:-|:-|
|`enabled`|On / off.|
|`graft_mask`|0 = kept exactly (default). 0.2-0.3 = kept roughly. 1 = ignored.|
|`source_clip`|The clip the keystone and capstone come from (1). On that shot and earlier nothing is grafted. Only used when `shot_number` is plugged in.|
|`placement`|`keystone front` (default), `keystone front + back`, `capstone front + back`, `keystone front + capstone back`. front = the shot's first latent, back = its last single-frame latent.|

|Input|Meaning|
|:-|:-|
|`clip 1 (keystone / capstone)`|A Load Latent fixed on clip 1. Keep it **outside** the Chain group so the Chain buttons don't change its number.|
|`shot_number`|From Clown Shot Counter. On shot 1 clip 1 isn't loaded (it doesn't exist yet).|

**Clown Late Lock** goes between Keystone and the sampler: latent from Keystone in, latent out to the sampler; model in, model out to the guider. For the first `swap_step` steps (default 4) the grafted slot is free, so the shot decides its own motion and pose. From the next step on it's locked to what Keystone put there. `slots` picks which slot: `capstone slot (back)` (default), `keystone slot (front)`, or both. It only touches slots Keystone actually filled, so on shot 1 it does nothing.

A capstone locked from the start pulls the pose back toward clip 1. Late Lock avoids that.

Comparison, clip 10 of the same 10-clip chain: no Keystone (base), keystone front, and keystone front + Late Lock capstone.

In a 10-clip test (measured about 20 frames from the end of clip 10), colour drift from clip 1 went from 6.4 with no Keystone, to 3.9 with keystone front, to 2.9 with keystone front + Late Lock capstone. Over-sharpening dropped as well.

**Joining:** the grafted frames are clip 1's, not part of the shot, so use the 17 + 22 cut: cut the last 17 frames of each clip and the first 22 of the next. In Seamless Video Combiner: **Use Beginning pair** on, Beginning Overlap 39, Beginning Join 22.

document fetched 2026-10-30 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xi7cng/try_this_workflow_please_it_cuts_degradation_in