Skip to content

Add missing CLIPLoader types and fp8_e4m3fn_fast weight dtype - #1801

Open
deepratna-awale wants to merge 1 commit into
LykosAI:mainfrom
deepratna-awale:add-missing-clip-loader-types
Open

deepratna-awale wants to merge 1 commit into
LykosAI:mainfrom
deepratna-awale:add-missing-clip-loader-types

Conversation

@deepratna-awale

@deepratna-awale deepratna-awale commented Nov 18, 2026 •

Copy link
Copy Markdown
Contributor

Fixes #1786

Problem

The Inference Type dropdown in the Text Encoders section is a hardcoded list that has fallen behind ComfyUI's CLIPLoader. krea2 is missing, so Krea 2 models can only be run with a wrong encoder type and fail with Krea2 expects conditioning with 12x2560 features. The same goes for every other single-encoder type ComfyUI has added since the list was last updated.

The model cards' Precision dropdown has the same problem: ComfyUI's UNETLoader offers fp8_e4m3fn_fast, but only the Wan model card listed it.

Changes

  • ModelCardViewModel.ClipTypes: added the CLIPLoader types that were missing compared to current ComfyUI: ovis, longcat_image, cogvideox, lens, pixeldit, ideogram4, boogu, krea2, joyimage, mage, minimax, yue2, flux3, moss_image, fibo2, pan2. The DualCLIPLoader types were already complete.
  • All new types are single-encoder types, so they fall through to the existing default of one encoder slot and are loaded with CLIPLoader / CLIPLoaderGGUF without further changes.
  • ModelCardViewModel.WeightDTypes and UnetModelCardViewModel.WeightDTypes: added fp8_e4m3fn_fast, matching WanModelCardViewModel and ComfyUI's UNETLoader.

Screenshots

The Inference model card, rendered headless on main and on this branch with the same steps: UNet loader, Custom workflow, then krea2 picked as the encoder type and fp8_e4m3fn_fast as the precision.

Before After

On main neither value is in its list, so both boxes stay empty.

Testing

  • New tests in ModelCardWorkflowProfileTests: the clip type list contains krea2 and other new single-encoder types and has no duplicates, and the weight dtype list offers fp8_e4m3fn_fast.
  • StabilityMatrix.Tests: all passed except the TestLicenseUrl / TestPreviewImageUri checks, which need outbound network access that my test environment blocks (unrelated to this change).

Sync the Inference encoder type dropdown with ComfyUI's CLIPLoader
options (krea2, ovis, longcat_image, flux3, ...) and offer the
fp8_e4m3fn_fast UNETLoader weight dtype in the model cards.

Fixes LykosAI#1786
@deepratna-awale
deepratna-awale marked this pull request as ready for review November 19, 2026 01:17

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Inference encoder Type dropdown missing krea2

1 participant