<h2 id="asian-fusion-krea-2-(v2-uncensored)"><a rel="ugc" href="https://latentmaster.com">Asian Fusion — Krea 2 (V2, Uncensored)</a></h2>Photoreal Asian faces and skin on Krea 2, tuned to look natural and candid instead of plastic. It covers the full range from clean portraits to fully uncensored, and it runs on the fast 8 step turbo base, so images come out quick.V2 is softer and calmer than V1, and it holds the Asian look without being pushed. Skin and light behave gently, faces stay understated, and it does not force explicit unless you ask it to.<h3 id="what-changed-since-v1">## What changed since V1</h3>V1 needed to be pushed hard before faces stopped drifting Western. V2 holds the Asian look on its own, and it is gentler with skin and light — less forced, more like a photo and less like a render. <a target="_blank" rel="ugc" href="https://latentmaster.com/?s=civ_LMAF">Get Premium models and updates</a><h3 id="recommended-settings">## Recommended settings</h3>Steps: 8 · CFG: 1 · Sampler: Euler · Scheduler: Simple · Size: 1024x1024 or larger.Text encoder: Qwen3-VL 4B (qwen3vl_4b_fp8_scaled). VAE: qwen_image_vae.At CFG 1 the negative prompt has no effect, so shape the image with the positive prompt.<h3 id="hardware-recommendations">## Hardware recommendations</h3>* 6 GB VRAM: the Q2_K GGUF. It is the only one that fits, and it is why it exists.* 8 GB VRAM: Q4_0 GGUF, or NVFP4 if you are on a 50-series card. Both will offload a little.* 12 GB VRAM: Q4_0 or NVFP4, comfortably.* 16 GB VRAM: INT8-ConvRot if you are on NVIDIA, otherwise fp8-scaled.* 24 GB VRAM and up: Q8_0 GGUF, the closest thing to the full model.Under 6 GB: it will still run — ComfyUI moves part of the model into system RAM — but it gets slow. Q2_K is the one to try.On an AMD card: Q8_0 GGUF first, fp8-scaled second. Skip INT8-ConvRot and NVFP4, those need NVIDIA. Q4_0 and Q2_K work fine on AMD too.Quality: from Q8_0 down to Q4_0 the difference is small and you will struggle to see it in a normal image. Take the biggest one that fits and stop worrying about it.About Q2_K: this one is squeezed hard, so it gets an honest paragraph. It is our own implementation — nothing available could write a Q2 for this architecture, so we wrote the quantizer ourselves and checked it byte for byte against ComfyUI's own reader. It holds up better than you would expect: poses, hands, groups and prompt accuracy all survive, and the extra grain suits this model. But it is not the same picture as the bigger files. The same prompt and seed give you a different image, not the same image slightly softer. It is also not faster — GGUF unpacks while it works, so Q2_K renders a little slower than NVFP4 despite being smaller. Take Q2_K because your card is small, not because you want speed.<h3></h3><h2 id="asian-fusion-krea-2-(v1-uncensored)">Asian Fusion — Krea 2 (V1, Uncensored)</h2>Photoreal Asian faces and skin on Krea 2, tuned to look natural and candid instead of plastic. It covers the full range from clean portraits to fully uncensored, and it runs on the fast 8 step turbo base, so images come out quick.Recommended settingsSteps: 8 · CFG: 1 · Sampler: Euler · Scheduler: Simple · Size: 1024x1024 or larger.Text encoder:Qwen3-VL 4B (qwen3vl_4b_fp8_scaled). VAE: qwen_image_vae.At CFG 1 the negative prompt has no effect, so shape the image with the positive prompt.<a target="_blank" rel="ugc" href="https://latentmaster.com/?s=civ_LMAF">Get Premium models and updates</a>

Asian Fusion Krea2 (Uncensored)

<a target="_blank" rel="ugc" href="https://latentmaster.com">Asian Fusion — Krea 2 (V2, Uncensored)</a>Photoreal Asian faces and skin on Krea 2, tuned to look natural and candid instead of plastic. It covers the full range from clean portraits to fully uncensored, and it runs on the fast 8 step turbo base, so images come out quick.V2 is softer and calmer than V1, and it holds the Asian look without being pushed. Skin and light behave gently, faces stay understated, and it does not force explicit unless you ask it to.## What changed since V1V1 needed to be pushed hard before faces stopped drifting Western. V2 holds the Asian look on its own, and it is gentler with skin and light — less forced, more like a photo and less like a render. <a target="_blank" rel="ugc" href="https://latentmaster.com/?s=civ_LMAF">Get Premium models and updates</a>## Recommended settingsSteps: 8 · CFG: 1 · Sampler: Euler · Scheduler: Simple · Size: 1024x1024 or larger.Text encoder: Qwen3-VL 4B (qwen3vl_4b_fp8_scaled). VAE: qwen_image_vae.At CFG 1 the negative prompt has no effect, so shape the image with the positive prompt.## Hardware recommendations* 6 GB VRAM: the Q2_K GGUF. It is the only one that fits, and it is why it exists.* 8 GB VRAM: Q4_0 GGUF, or NVFP4 if you are on a 50-series card. Both will offload a little.* 12 GB VRAM: Q4_0 or NVFP4, comfortably.* 16 GB VRAM: INT8-ConvRot if you are on NVIDIA, otherwise fp8-scaled.* 24 GB VRAM and up: Q8_0 GGUF, the closest thing to the full model.Under 6 GB: it will still run — ComfyUI moves part of the model into system RAM — but it gets slow. Q2_K is the one to try.On an AMD card: Q8_0 GGUF first, fp8-scaled second. Skip INT8-ConvRot and NVFP4, those need NVIDIA. Q4_0 and Q2_K work fine on AMD too.Quality: from Q8_0 down to Q4_0 the difference is small and you will struggle to see it in a normal image. Take the biggest one that fits and stop worrying about it.About Q2_K: this one is squeezed hard, so it gets an honest paragraph. It is our own implementation — nothing available could write a Q2 for this architecture, so we wrote the quantizer ourselves and checked it byte for byte against ComfyUI's own reader. It holds up better than you would expect: poses, hands, groups and prompt accuracy all survive, and the extra grain suits this model. But it is not the same picture as the bigger files. The same prompt and seed give you a different image, not the same image slightly softer. It is also not faster — GGUF unpacks while it works, so Q2_K renders a little slower than NVFP4 despite being smaller. Take Q2_K because your card is small, not because you want speed.

Files with SHA256: ad10b8117e8769f42ee2955bd6b8c9020387a8dc594326c707b60cab9b871b29