Add unified set of Diffusion model transforms electric boogaloo #57
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
"Once I add velocity this can trivially go into the diffusers wrapper" yeah pretty sure it's actually impossible to write v-prediction as z̃ₛ = z̃ₜ + output · (ηₛ - ηₜ) so I rewrote it again into a custom parameterization space that actually for real works with everything and am wiping out all of the commits trying to make DiffusionFlow work Squashed commit of the following: commit 6b83bf297e87b888a65733e1ed0c95b30f21bbd0 Author: Beinsezii <[email protected]> Date: Mon Nov 17 01:11:55 2025 -0800 Rewrite ModelTransform to use custom parameterization commit 4913744 Author: Beinsezii <[email protected]> Date: Sun Nov 16 16:45:17 2025 -0800 Velocity yes depend on X commit 4373aa1 Author: Beinsezii <[email protected]> Date: Fri Nov 14 22:56:11 2025 -0800 Velocity don't depend on Epsilon commit 0d0a4c0 Author: Beinsezii <[email protected]> Date: Fri Nov 14 22:22:55 2025 -0800 Fix FlowModel to_x/from_x with VariancePreserving schedule commit d6b5971 Author: Beinsezii <[email protected]> Date: Fri Nov 14 03:49:50 2025 -0800 Make DiffusionModel the X impl commit 73a73d6 Author: Beinsezii <[email protected]> Date: Fri Nov 14 03:22:07 2025 -0800 models: Add XModel, VelocityModel, .to_x(), .from_x(), .to_h() commit 26b012a Author: Beinsezii <[email protected]> Date: Tue Nov 11 01:38:37 2025 -0800 Add unified set of Diffusion model transforms Currently just for sampling.functional Once I add velocity this can trivially go into the diffusers wrapper instead of a raw prediction function, and then moved through structured Possibly I could associate the schedule and therefore also the sigma transform `type base_schedule = Linear` or whatever. Will have to play around see if FlowModel can be made to handle a Variance-Preserving schedule and vice-versa. Just so fucking glad I don't have to switch on sigma_transform anymore for the tableau solver.
Still not 100% certain on how to best parameterize the derivative transform. It needs to be per-step for plans on making RKUltra its own diffusers wrapper, so ig adding yet another parameter™ is hot.
Since we have common.divf() now this is more appropriate
…atamodel,NoiseModel
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
"Once I add velocity this can trivially go into the diffusers wrapper"
yeah pretty sure it's actually impossible to write v-prediction as
z̃ₛ = z̃ₜ + output · (ηₛ - ηₜ)
so I rewrote it again into a custom parameterization space that actually for real works with everything and am wiping out all of the commits trying to make DiffusionFlow work
replaces #55