text-to-image diffusion model
based on original Krea2 Raw model
blended with a bunch of loras
not baked with turbo lora
NSFW available but not extreme
Description
FAQ
Comments (7)
Do you have bigger version? (bf16, fp8 or int8)
yes i have the bf16, i will upload it asap
bf16 is available
Gives error on Forge:
RuntimeError: Error(s) in loading state_dict for SingleStreamDiT: size mismatch for blocks.0.attn.wq.weight: copying a param with shape torch.Size([6144, 3072]) from checkpoint, the shape in current model is torch.Size([6144, 6144]). size mismatch for blocks.0.attn.wk.weight: copying a param with shape torch.Size([1536, 3072]) from checkpoint, the shape in current model is torch.Size([1536, 6144]). size mismatch for blocks.0.attn.wv.weight: copying a param with shape torch.Size([1536, 3072]) from checkpoint, the shape in current model is torch.Size([1536, 6144]). (etc., the full output is impossible long)
Other checkpoints typically work in NVFP4. Redownloaded couple times, the same error.
ok i don't know the reason about your issue.
i don't use Forge either.
may be it's due to my usage of the kitchen converter.
sorry. i'll try to fix it asap
Ok but if you don't publish an FP8 or int8, and a Turbo version then you'll get like 1 in 10 who bother to even try it, probably the newbs too. NVFP4 is visibly lower quality than those. There's about a dozen new models posted in the past two days in int8, fp8 and turbo.
There are many people who has weak cards and 8-bit checkpoint make their GPUs suffer. Unfortunately KREA has rather heavy "vision" LLM TE which gives additional problems on 8gigs.