【R9700/RX7900XT】Comfy UI 的 ck-attention 配合8步 加速 LoRa 效果明显，强烈推荐大家试一下。
-
如题，Comfyui 现支持 CK Attention 和 Rocm 动态显存，Minimax H3 总完成时间与基线缩短一半。
我的配置是CPU：R7 9700X，内存：DDR5-6000 2x24GB，显卡：R9700；系统Windows 11。
ComfyUI 版本：0.33.1 & ROCm 7.14 & PyTorch 2.12.0，添加以下启动项后：
--use-ck-attention --enable-dynamic-vram生成时间为 180.02秒（8/8 [01:56<00:00, 14.58s/it] ），而不添加此启动项则为 284.44秒（8/8 [03:14<00:00, 24.28s/it]）秒，约为1.6倍，效果显著。
我测试的视频随机种子为 43 ；视频比例为 16：9；视频长度为 8秒；0.4MP的清晰度；
这是我的测试提示词：
integrated_multimodal_description: [Shot 1] Live-action, cinematic, dreamlike photorealism. At blue hour beneath a deep indigo sky, a young East Asian woman with long black hair walks slowly through a vast field of delicate white flowers covered by ankle-high silver mist. She wears a flowing translucent white chiffon dress with subtle pearl embroidery. A low-angle tracking shot follows behind her with smooth, gentle movement as her hair, sleeves, and layered skirt drift naturally in the soft night breeze. Cool moonlight creates realistic highlights along the fabric and flower petals, while faint warm light glows beneath the mist. A luminous blue butterfly enters from the right, circles once around her shoulder, and settles lightly on her extended fingertip. She pauses and turns into a graceful three-quarter profile, watching it with a subtle, peaceful expression. The camera begins a slow arc around her with small amplitude. The butterfly dissolves into hundreds of tiny blue-white particles that flow from her hand like sparkling dust. A circular wave of light travels outward through the field, causing the flowers to illuminate one after another in a smooth expanding pattern. The camera gently cranes upward and moves past her, revealing a perfectly still mirror-like lake beyond the flowers and an enormous luminous full moon resting close to the horizon. The moon, stars, glowing flower field, and drifting particles reflect naturally in the water. The final frame holds briefly on the woman standing alone before the moonlit lake as several glowing particles float upward. Realistic human anatomy, consistent facial identity, physically accurate fabric and hair movement, volumetric moonlight, shallow depth of field, subtle lens bloom, fine film grain, high dynamic range, elegant restrained color grading, no cuts, no dialogue, no visible text. overall_soundscape: A soft night breeze moves through the flowers and gently rustles the woman's chiffon dress. Her quiet footsteps brush against the grass, accompanied by a delicate crystalline shimmer when the butterfly dissolves and the flowers begin to glow. non_diegetic_music: Sparse felt-piano notes at a slow tempo, joined by a sustained ethereal string pad and very light glass harmonics. The music gradually rises as the flowers illuminate, then resolves into a soft sustained note over the final frame.另外也可以通过安装节点来进一步加速（不过我嫌麻烦没有折腾），节点链接：
https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3
Reddit原帖：
https://www.reddit.com/r/comfyui/comments/1vo9zjj/comfyui_now_supports_ck_attention_and_dynamic/
另外还有个利用Ltx2.5为MiniMax H3 放大的放帖子我觉得也不错，所以也一并搬来了，但我没有尝试：
-
，
T terry 固定了此主题
-
@Bunsei 这个"快是快了，效果一拖"值得拆开看，别把整套提速方案一棍子打死：
- ck-attention 单独开基本是数值安全的——它只是把 attention 换成 Triton 优化的 kernel，算法和原版一致，180s vs 284s 那档对比纯粹是速度收益，一般不会改变出图结果。
- 真正的质量嫌疑在 Spectrum（sage-attention 类节点）：它把 Q/K 量化到 int8、V 量化到 fp8 做有损 attention。对 H3 这种长序列视频模型，量化误差会直接反映到画面细节和一致性上——143s 那档最猛的速度，代价大概率就付在这里。
- 隔离验证最快：同一 prompt，三个开关分别单独开——①只开 ck-attention ②只开 Spectrum ③两个都开。如果①的出图回到基线，问题就锁定在 Spectrum，ck-attention + dynamic-vram 可以放心保留。
- 另外 8 步 Turbo LoRA 本身步数就少、容错低，再叠加有损 attention，误差会被进一步放大，"一拖"会更明显。
建议：出图质量优先就只留 ck-attention（还是能省 284→180s 那档）；速度优先且能接受细节损失，再开 Spectrum。
-
，系统 取消固定了此主题
