<h3 id="meet-dogepoint-the-first-model-for-sensitive-young-mentm!">Meet DogePoint, the first model for sensitive young men™!</h3>To answer the "wat checkpoint?" question on all of my showcases, these EPS and VPred bases are available so you can reproduce everything 1:1 or better, there are no other specialized models for anime, 3D or realism, these models should cover it all.These models are calibrated around Euler A, CFG 5, Normal/SGM Uniform schedulers and resolutions around 1280 and 1536, you can go beyond that but the fracture risk, quality increase and rendering times are just not worth it. The showcased images are latent upscaled but you can generate straight on the final resolution without significant loss, which I recommend if you're not familiar with upscaling configurations and implications that would need an entire article to list part of it.The main goal here is to run LoRA/LoCon with the best quality and the least interference possible, so I do recommend running mainly Illustrious and NoobAI trained models and avoid everything else, even finetunes labeled as IL that might mess up your blocks, give you blurry and washed up images.V1 UPDATE: So it's here and it's chopped, quite literally. <s>Since civit now supports it and people are used to having separated files, I won't upload two large models just for the pruned text encoder. If you have more than 8GB VRAM go for the TE32 version, specially if you intend to use hires-fix and latent upscaling</s>. <s>The current VAE weights are scuffed on some settings and you can probably skip it until the next update</s>. As expected, the showcased images are generated without adetailer, no extra script, default settings and with reasonable minimum sampling and CPU for RNG source, you can get better results and higher LoRA fidelity by going for GPU, for example. *Uploaded the previous candidate (V0) UNET in case you're having dimension issues and a TE with both headers so you can make them work virtually anywhere, as for the V1, it's up again without the VAE issues and the same old TE that might show issues on some platforms, remap it with diffusers, "unstrict" your state_dict, change platforms, wait for V2 or whatever, but changing the files from the showcase kinda defeats its purpose. It's noteworthy to mention that the alpha models aren't 100% finished, some improvements on sharpness, color temperature and small changes are still needed to clean and wrap it up. Image feedback is appreciated and will be taken into account for improvements. Any suggestions are welcome, let me know if there's anything else you need uploaded or written like the old "legacy base" and an upscaling guide.

DogePoint - NAI

dogepointNAI_dogenaiEPSV1_3292326.safetensors

dogepointNAI_dogenaiEPSV1_3305019.safetensors

dogepointNAI_dogenaiEPSV1_txt_fp16.safetensors

dogepointNAI_dogenaiEPSV1_txt_fp32.safetensors

dogepointNAI_dogenaiEPSV1_3304679.safetensors

DogeNAI_V1-VAE.safetensors

dogepointNAI_dogenaiEPSV1_3297086.safetensors

dogepointNAI_dogenaiEPSV1_fp16.safetensors

dogepointNAI_dogenaiEPSV1_3305018.safetensors

dogepointNAI_dogenaiEPSV1_txt.safetensors

DogeNAI - EPS V1

DogeNAI_BETA - EPS

DogeNAI_ALPHA - VPred

Files with SHA256: 86d2e8acba932aebab047a82beebc54729b87a8f679fdc94c780887c29693274