Qwen-Image Sharp VAE [Anima / Krea-2]
TL;DR
EN
- A sharper, clearer drop-in replacement for the Qwen-Image VAE: crisper lines, eyes and hair, with no overexposure.
- 3 strengths (pick from the download dropdown): x0.5 (subtle) / Default (x1.0) / x1.5 (strongest).
- Works with Anima, Krea-2, Qwen-Image.
中文
- 比原版更锐利、更清晰的 Qwen-Image VAE，直接替换就能用。线条、眼睛和发丝更清楚，而且不会过曝。
- 3 个强度（在下载按钮旁边的下拉菜单里选）：x0.5（轻度）/ Default（x1.0）/ x1.5（最强）。
- 适用于 Anima、Krea-2、Qwen-Image。
Crisper lines, eyelashes and hair strands, just by swapping the VAE. The difference is biggest with hires fix.
A drop-in replacement for qwen_image_vae for Anima and Krea-2. No trigger words, no extra nodes, no retraining. Same colors, same exposure, sharper details.
Same prompt, same seed, same settings. Only the VAE is different.
| Scene | Original: sharpness / clipped | Qwen-Image Sharp: sharpness / clipped | Sharpness gain |
|---|---|---|---|
| Angel ¹ | 6.48 / 3.04% | 11.26 / 1.60% | +74% |
| Miku | 6.87 / 4.31% | 11.21 / 1.82% | +63% |
| Portrait | 3.89 / 19.89% | 7.65 / 17.49% | +97% |
| Lineart classroom | 6.21 / 15.75% | 9.01 / 14.33% | +45% |
| Backlit beach | 2.45 / 8.70% | 3.67 / 8.26% | +50% |
| Castle scenery | 3.81 / 1.80% | 5.78 / 0.05% | +52% |
| Summer festival | 8.72 / 0.94% | 14.08 / 0.70% | +61% |
Every scene gets sharper, and every scene also has fewer blown-out highlights than the original.
- Sharpness = mean luminance gradient (higher = crisper edges).
- Clipped = share of pixels with any channel ≥ 254.
- Settings: SilvermoonMix-Anima29B-Evolved-V2.3, Forge Neo, Euler a, Simple, 28 steps, CFG 3, Shift 3, 832×1216. Hires fix: 1.5×,
4x-AnimeSharp, denoise 0.15, 8 hires steps, hires CFG 3, hires Shift 3. - ¹ Angel: same settings, but 14 hires steps and ADetailer (face_yolov8n, denoise 0.15).
Why the VAE matters so much in hires fix
In hires fix, the VAE is not only used at the very end. It also sits in the middle of the pipeline:
- The first pass produces a small latent.
- The VAE decodes it into an image. ← Qwen-Image Sharp gives a crisper image here.
- The upscaler (e.g.
4x-AnimeSharp) enlarges that image. Clean edges in, clean edges out. - The enlarged image is encoded back into a latent and the model resamples it at the chosen denoise strength.
- The VAE decodes the final image.
So the VAE decides what the second pass starts from. With a low denoise strength, the second pass mostly keeps that starting image, and the extra detail from step 2 carries all the way to the final result. That is why only swapping the VAE makes such a large difference in hires fix.
Effect of denoise strength
Same seven scenes, hires fix with denoise 0.15 / 0.30 / 0.45 (everything else unchanged):
| Scene | 0.15 | 0.30 | 0.45 |
|---|---|---|---|
| Angel ¹ | +74% | +63% | +53% |
| Miku | +63% | +56% | +46% |
| Portrait | +97% | +88% | +78% |
| Lineart classroom | +45% | +39% | +36% |
| Backlit beach | +50% | +40% | +31% |
| Castle scenery | +52% | +37% | +26% |
| Summer festival | +61% | +53% | +42% |
(Sharpness gain of Qwen-Image Sharp over the original VAE at the same denoise.)
- Lower denoise keeps more of what the VAE decoded, so the gain is largest at 0.15.
- Higher denoise lets the model redraw more, so the gap narrows, but it is still +26% to +78% at 0.45.
- Qwen-Image Sharp had fewer clipped highlights than the original in all 21 scene/denoise combinations.
Not your taste? Try x0.5 or x1.5
If the default feels too sharp or not sharp enough, two more strengths are included:
| File | Strength | Good for | SHA256 |
|---|---|---|---|
qwen_image_sharp_vae_x0.5_bf16.safetensors |
×0.5 | Soft, painterly or thick-paint styles | 891b35f59f1fa08d129e0de406421312cab94050d3b00dcc4415c835aec64936 |
qwen_image_sharp_vae_bf16.safetensors |
default | Recommended for most use | 806fa84dd9a6b29b8e9bfce64804c51a9bc71f44f2b03ab36b1d2c1a05b4f154 |
qwen_image_sharp_vae_x1.5_bf16.safetensors |
×1.5 | Very clean cel shading / lineart, if you want maximum crispness | 0cf7faf49a5544a44978dce1b6e39c88567e62c957424c7f1ade1309ff4a089e |
All files are 242 MB (BF16).
Sharpness in hires fix (denoise 0.15):
| Scene | Original | x0.5 | Default | x1.5 |
|---|---|---|---|---|
| Miku | 6.87 | 8.82 | 11.21 | 13.58 |
| Portrait | 3.89 | 5.54 | 7.65 | 9.72 |
| Backlit beach | 2.45 | 2.88 | 3.67 | 4.63 |
| Castle scenery | 3.81 | 4.17 | 5.78 | 8.07 |
| Summer festival | 8.72 | 10.80 | 14.08 | 17.54 |
About x1.5: in hires fix it adds heavy dark outlines and can leave light halos around edges (see the enlarged crops above). In the festival scene it was the only variant with more clipped highlights than the original (1.17% vs 0.94%). Use it only if you really want the strongest effect.
More: strength_02_portrait, strength_04_backlight, strength_05_scenery. Every single image with full generation parameters is in images/samples.
Usage
- Forge Neo: put the file in
models/VAEand restart the WebUI (refresh does not register new files). Select it in the VAE / Text Encoder module list together withqwen_3_06b_base. For hires fix, keep Hires Module 1 on Use same choices. - ComfyUI: put the file in
models/vaeand load it withLoad VAE. In a hires workflow, use it for the VAE Decode before the upscale as well as the final decode.
FAQ
Does it affect LoRA training? No. The encoder is byte-identical to the original Qwen-Image VAE, so training latents are exactly the same. Use it for generation only.
Does it work with other models? It works with models that use the Qwen-Image VAE (qwen_image_vae.safetensors). Krea-2 ships the exact same VAE file (identical SHA256), and Qwen-Image / Qwen-Image-Edit use the same weights. It was tuned and tested on Anima; image quality on Krea-2 and Qwen-Image was not tested. It does not work with Z-Image (Flux VAE) or MOSS-Image (Flux 2 VAE).
Latent upscaler? A latent upscaler skips the decode in the middle of hires fix, so the gain is expected to be much smaller. All tests here used a pixel upscaler (4x-AnimeSharp).
How it was made
No training. The decoder was rebuilt from the weight difference between the original Qwen-Image VAE and Felldude's Qwen-Image HDR VAE decoder:
- For each decoder convolution, the difference was split into a zero-mean edge part (sharpening) and a mean part (brightness / color shift).
- The edge part was kept in all layers. The brightness part was removed in the low-resolution layers and only lightly kept in the full-resolution layers for darker line ink. This keeps the original colors and exposure.
decoder.head, all biases and all norm weights are unchanged from the original.- x0.5 and x1.5 scale the whole change. All math in FP32, rounded to BF16 once.
Credits & License
- Base VAE: Qwen/Qwen-Image, Apache-2.0.
- Decoder weight difference: Felldude/Qwen-Image-HDR-VAE, Apache-2.0.
- This release modifies the decoder weights as described above and is distributed under Apache-2.0.
Qwen-Image Sharp VAE [Anima / Krea-2]（中文说明）
只换一个 VAE，线条、睫毛、发丝都更清晰。开高清修复时效果最明显。
可直接替换 Anima 和 Krea-2 的 qwen_image_vae，不需要触发词、额外节点或重新训练。色彩和曝光保持原版，细节更锐利。
同提示词、同种子、同参数，只换 VAE（高清修复 1.5 倍，重绘幅度 0.15）：
| 场景 | 原版 锐度 / 溢出 | Qwen-Image Sharp 锐度 / 溢出 | 锐度提升 |
|---|---|---|---|
| 天使 ¹ | 6.48 / 3.04% | 11.26 / 1.60% | +74% |
| 初音 | 6.87 / 4.31% | 11.21 / 1.82% | +63% |
| 人像特写 | 3.89 / 19.89% | 7.65 / 17.49% | +97% |
| 教室线稿 | 6.21 / 15.75% | 9.01 / 14.33% | +45% |
| 海边逆光 | 2.45 / 8.70% | 3.67 / 8.26% | +50% |
| 城堡风景 | 3.81 / 1.80% | 5.78 / 0.05% | +52% |
| 夏祭双人 | 8.72 / 0.94% | 14.08 / 0.70% | +61% |
每个场景都更锐利，高光溢出也都比原版少。
参数：SilvermoonMix-Anima29B-Evolved-V2.3，Forge Neo，Euler a，Simple，28 步，CFG 3，Shift 3，832×1216；高清修复 1.5 倍，4x-AnimeSharp，重绘幅度 0.15，Hires 8 步，Hires CFG 3，Hires Shift 3。
¹ 天使：参数相同，但 Hires 14 步，并开启 ADetailer（face_yolov8n，重绘幅度 0.15）。
为什么 VAE 对高清修复影响这么大
高清修复时，VAE 不只在最后用一次，中间也会用到：
- 第一遍生成一张小尺寸 latent。
- VAE 把它解码成图片。Qwen-Image Sharp 在这一步就给出更清晰的图。
- 放大器（如
4x-AnimeSharp）放大这张图。输入的边缘干净，放大后的边缘也干净。 - 放大后的图编码回 latent，模型按重绘幅度重新采样。
- VAE 解码出最终图片。
所以第二遍从什么样的图开始画，是由 VAE 决定的。重绘幅度低时，第二遍基本保留这张起始图，第 2 步多出来的细节会一直带到最终结果。这就是只换 VAE、高清修复结果就差很多的原因。
不同重绘幅度的影响
Qwen-Image Sharp 相对原版的锐度提升：
| 场景 | 0.15 | 0.30 | 0.45 |
|---|---|---|---|
| 天使 ¹ | +74% | +63% | +53% |
| 初音 | +63% | +56% | +46% |
| 人像特写 | +97% | +88% | +78% |
| 教室线稿 | +45% | +39% | +36% |
| 海边逆光 | +50% | +40% | +31% |
| 城堡风景 | +52% | +37% | +26% |
| 夏祭双人 | +61% | +53% | +42% |
- 重绘幅度越低，VAE 解码出的细节保留得越多，0.15 时提升最大。
- 重绘幅度越高，模型重画的部分越多，差距会缩小，但到 0.45 仍有 +26%～+78%。
- 21 组组合里，Qwen-Image Sharp 的高光溢出全部低于原版。
觉得太锐或不够锐？还有 x0.5 和 x1.5
| 文件 | 强度 | 适合 |
|---|---|---|
qwen_image_sharp_vae_x0.5_bf16.safetensors |
×0.5 | 厚涂、柔光画风 |
qwen_image_sharp_vae_bf16.safetensors |
默认 | 大多数情况推荐 |
qwen_image_sharp_vae_x1.5_bf16.safetensors |
×1.5 | 想要最锐利的赛璐璐、线稿风 |
高清修复（重绘幅度 0.15）下的锐度：
| 场景 | 原版 | x0.5 | 默认 | x1.5 |
|---|---|---|---|---|
| 初音 | 6.87 | 8.82 | 11.21 | 13.58 |
| 人像特写 | 3.89 | 5.54 | 7.65 | 9.72 |
| 海边逆光 | 2.45 | 2.88 | 3.67 | 4.63 |
| 城堡风景 | 3.81 | 4.17 | 5.78 | 8.07 |
| 夏祭双人 | 8.72 | 10.80 | 14.08 | 17.54 |
关于 x1.5：开高清修复时描边会明显变粗变黑，边缘可能出现亮边（见上图放大区域）。夏祭场景里，它是唯一一个溢出比原版高的版本（1.17% vs 0.94%）。确实想要最强效果时再用。
使用方法
- Forge Neo：放进
models/VAE后重启 WebUI（点刷新不会登记新文件），在 VAE / Text Encoder 模块里和qwen_3_06b_base一起选上；高清修复的 Hires Module 1 保持 Use same choices。 - ComfyUI：放进
models/vae，用Load VAE加载。高清修复工作流里，放大前的那次 VAE Decode 和最终解码都要用它。
常见问题
会影响 LoRA 训练吗？ 不会。编码器与原版 Qwen-Image VAE 逐字节相同，训练用的 latent 完全一样，只用于出图即可。
能用在其他模型上吗？ 使用 Qwen-Image VAE（qwen_image_vae.safetensors）的模型都能加载。Krea-2 自带的 VAE 与原版是同一个文件（SHA256 相同），Qwen-Image / Qwen-Image-Edit 的 VAE 权重也相同。参数是在 Anima 上调整和测试的，Krea-2 和 Qwen-Image 的画面效果没有测试过。Z-Image（Flux VAE）和 MOSS-Image（Flux 2 VAE）用的是不同的 VAE，不能使用。
用 Latent 放大器有效吗？ Latent 放大器会跳过高清修复中间那次解码，预计提升会小很多。本页所有测试用的都是像素放大器 4x-AnimeSharp。
制作方法
没有经过训练，而是用原版 Qwen-Image VAE 和 Felldude 的 Qwen-Image HDR VAE 解码器之间的权重差值重新组合：
- 每个解码器卷积的差值拆成两部分：零均值的边缘部分（锐化）和平均值部分（亮度、色偏）。
- 所有层保留边缘部分；低分辨率层去掉亮度部分，全分辨率层只少量保留，用来加深线条墨色。这样色彩和曝光保持原版。
decoder.head、所有 bias 和 norm 权重与原版相同。- x0.5 / x1.5 按比例缩放整体改动。全程 FP32 计算，最后只做一次 BF16 舍入。
致谢与许可
- 原版 VAE：Qwen/Qwen-Image，Apache-2.0
- 解码器权重差值来源：Felldude/Qwen-Image-HDR-VAE，Apache-2.0
- 本发布修改了解码器权重（见“制作方法”），以 Apache-2.0 许可发布。