Type | |
Stats | 203 |
Reviews | (36) |
Published | Oct 20, 2024 |
Base Model | |
Hash | AutoV2 E258DB2938 |
This is basically a stabilization of the v-prediction test of NoobAI. Weights from IterComp XL were added onto the model using this method, then the original clip was added back in.
Perpendicular component is used with add difference to prevent interference with v-prediction noise.
Nodes can be found here (requires sd_mecha installed): https://github.com/ljleb/comfy-mecha
Iterative Composition-Aware Feedback Learning: https://arxiv.org/pdf/2410.07171
V-prediction support has been added to Automatic1111 recently. On ComfyUI, use the ModelSamplingDiscrete node.
Discussion
I'm impressed at how well this merge "stabilizes" the vpred test provided by the Noob team. Makes it quite usable and I've found it's quickly become my daily driver model over their current epsilon-pred release.
This is actually a really good improvement on the v-pred version base and the stabilizing effect is quite noticeable. In particular it stops the excessively wide (and often incorrect) swings in rendering style that I find the base noobai v-pred version has (though that might be because it too is undertrained somewhat still).
I hope you continue to update this version if they release newer noobai versions!
