Downloads · 30 days
77
5% of all-time downloads
AlekseyCalvin/Custom_VAE-Z-image-FLUX.1-by-G-REPA
Custom_VAE-Z-image-FLUX.1-by-G-REPA is a machine learning model from AlekseyCalvin. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for diffusers. The card lists the license as apache-2.0.
An experimental FLUX-based semantically-aligned VAE (Visual Auto Encoder) shared in July, 2025 by G-REPA. <br Original repo/weights found HERE. <br This checkpoint was drawn, in a rather adhoc way, from a vast pool of…
Downloads · 30 days
77
5% of all-time downloads
All-time downloads
1.4K
Public
Parameters
83.8M
335 MB on disk
Likes
2
Public
Click a slice to open those files.
.safetensors335 MB · 100%
From the Hugging Face model README
An experimental FLUX-based semantically-aligned VAE (Visual Auto Encoder) shared in July, 2025 by G-REPA. <br> Original repo/weights found HERE. <br> This checkpoint was drawn, in a rather adhoc way, from a vast pool of WIP G-REPA repo variants (now repeat that aloud thrice... faster!). <br> Its parent repo bears the catchy name of "G-REPA/Self-Attention-W2048-3B-Res256-VAEFLUX-Repa0.5-Depth8-Dinov2-B_100000". <br> This VAE may be used with FLUX.1, Z-image Turbo, & other FLUX.1 VAE-leveraging DiT models. It was confirmed to work in a plug & play way via Diffusers/ZeroGPU Spaces &, thus, probably locally as well. <br> It is producing decent results in my Z-Image Turbo LoRA the Explorer-type gallery space, though I've yet to perform reliable comparison testing. <br>
G-REPA is, as it seems, a WIP checkpoint archive for REPA-E. <br> In its turn, 'REPA-E' reps & repo's an international co-laboration exploring, actualizing, & furthering the potentials of representation-aligning (aka, uhmmm, 'RepA'-ing) VAEs in tandem w/ trad T2I LDm training to convergently improve both. <br> This quest, reportedly quite fruitful, extends their earlier/concurrent research in using a visual foundation encoder (DINOv2) to semantically-align the tokenization prowess of pretrained VAEs. <br> And it is from this semantic alignment track that the checkpoint herein derives. <br>
All credit goes out to the creators. <br> The only reason for my repost is to facilitate simpler Diffusers/Spaces pipelining. <br>
In Russian, 'repa' or (in Cyrillic) 'репа' means 'turnip', the root vegetable with a nutritious tuber and overground sprouts, and one of the oldest agricultural crops, cultivated as early as forty centuries ago in ancient Egypt, Greece, India, and across much of central Asia. <br> Throughout human history, the turnip had remained one of the most effectively and easily farmable edible crops and, consequently, one of the cheapest/most widely accessible food sources sustaining population expansions and/or mediating food scarcity. <br> Pior to the popularization of the potatoe, another root tuber introduced from the Americas, the turnip was the core staple of the worker's and the pauper's plate across much of the world. <br> Between the 17th and the 19th centuries, the turnip played a key role in farmland-optimization methodologies engendering manyfold agricultural output growth often called the "Second Agricultural Revolution". <br> Indirect effects of such agricultural innovations indirectly enabled and fostered industrial development across much of the world, by expanding, freeing up from the fields, and feeding the populations which came to constitute the industrial proletariat. <br>
"The moon like a turnip, and stars are beans; so, / I thank you, my mother, for bread and for salt..." <br> – Yegor Letov (in my translation)
In Tianjin, China, near the Tianjin Foreign Language University, an elaborate sculptural installation depicts kids pulling a giant turnip from the ground, though it sometimes gets mistaken for a carrot...