凤希AI - SeedVR2超分
The whole SeedVR2 chain in one node, with the memory bill up front
- 输入帧
- MODEL
- VAE
- 处理后帧
- 处理报告
- 放大倍数
SeedVR2, minus the five-node blueprint
SeedVR2 is ByteDance's one-step video restoration model that the community quietly turned into the default detail-restoration upscaler - Apache 2.0, and it still runs a 3x image upscale on a 6GB card. It's the model you reach for when you want detail rebuilt rather than interpolated.
The official ComfyUI blueprint wires it as five nodes - pre-process, tiled VAE encode, conditioning, sampler, post-processing - and you get to compute the 4n+1 frame padding, decide the tile sizes, and stitch the pieces back together yourself. FxAiSeedVR2Upscale (凤希AI - SeedVR2超分) is that whole chain in one node, plus the bookkeeping nobody wants to do: chunking, cross-fades, colour matching, and an honest estimate of what it's about to cost you.
How it works
The pipeline per block is: lanczos pre-scale → pad to 4n+1 → tiled VAE encode → one-step sample → tiled decode → colour transfer → crop back to the original frame count.
The first step is the one people get backwards. SeedVR2 doesn't scale. It reconstructs at a given size, so this node resizes with lanczos to your target dimensions first and lets the model rebuild detail there. So "4x" is really "invent detail at 4x".
Chunking is by pixel frames, not latent frames: a block goes in, encodes, samples, decodes and colour-matches, gets written out, and is freed. Peak memory is therefore a function of block size, not clip length - which is the entire reason this works on a consumer card. Adjacent blocks overlap by 8 frames by default and are blended with a Hann curve, and when a block OOMs the node halves it, retries, and remembers the smaller size for the rest of the run.
The colour pass (lab by default) pulls the generated frame's colour back toward the pre-scale reference.
The inputs you actually touch
- 输入帧 - single image or a frame sequence. Everything downstream is batched.
- 模型 - whatever
models/diffusion_modelsfile has "seedvr" in its name. Defaults to the 7B if one exists, which is the better pick: the author measured 6–7s per 1280×768 frame on 7B vs 3.7s on 3B, and in batch they're nearly identical because VAE tiling dominates. - 放大倍数 - 1.0 to 4.0, default 2. Cost scales with the square of this. 2x is the value sweet spot; 3x and 4x are paying for invention.
- 目标短边 - set it (720/1080/1440) to ignore the multiplier and normalise everything to a short edge. Dimensions snap to multiples of 16, so 1080 comes out 1088.
- 分块帧数 -
自动picks the least-wasteful 4n+1 in the 21–41 range for your actual frame count; bigger blocks are faster because the overlap is amortised. Drop to 9 or 13 if VRAM is tight. - 块间重叠帧, 色彩校正, 显存自适应, 随机种子 - leave defaults until something's wrong. The seed barely matters here (one-step sampling, strongly constrained by the input latent), which the tooltip admits outright.
Optional MODEL / VAE sockets let you feed a pre-loaded pair; otherwise it caches both internally and looks for a seedvr-named VAE under models/vae.
Outputs: 处理后帧 (IMAGE), 处理报告 (STRING - the timing and memory breakdown, genuinely worth reading once), and 放大倍数 (FLOAT, the scale actually reached). Wire that third one into the face-refine node's 放大倍数 input in this pack and its face-width thresholds scale themselves.
Installing
cd ComfyUI/custom_nodes
git clone https://github.com/fxai666/fxai-toolkit
Then the model side, which is on you - the pack ships no weights:
- a SeedVR2
.safetensorsinComfyUI/models/diffusion_models/with seedvr somewhere in the filename (the pack's reference workflow usesseedvr2_3b_int8_convrot.safetensors) - the matching SeedVR2 VAE in
ComfyUI/models/vae/, filename containingseedvr
And a ComfyUI build new enough to have comfy.ldm.seedvr - this node imports lab_color_transfer and friends from core, so an old install fails at import and the node simply won't appear.
Where it bites
The dropdown says no weights found. The filename filter didn't match; rename the file so "seedvr" appears in it.
It refuses to start, on purpose. Before running, it estimates the whole chain's memory - output frames plus a second copy for downstream face refining - and if that exceeds 90% of available RAM it stops with three suggestions: shorten the segment, lower the multiplier, or drop the face-refine stage. Better than the alternative, which is your machine swapping for 40 minutes and then dying. It needs psutil, which the pack installs on import.
Plasticky faces. SeedVR2 redraws the whole frame, so identity drifts - the node's own description puts it at 0.76–0.96 similarity. The pack's answer is to chain its face-refine node after this one, which only replaces pixels inside the face box and holds identity above 0.82. If your portraits look like mannequins, that's the fix, not a different upscaler.
Quantisation matters more than anything else here. The community verdict is blunt: FP16 is clean, FP8 adds tiling grids, GGUF damages skin. The dropdown will happily list your int8 and fp8 quants. If the output looks bad, check which file you loaded before blaming the node.
It's not fast. Batch the segment and walk away; it prints a minutes estimate for a reason.
Inputs (11)
| Name | Type | Default | Description |
|---|---|---|---|
| 输入帧 | IMAGE | 要超分的画面。接图像序列(视频帧列表)也行,单张也行。 会先把整段按「分块帧数」切开,一块一块跑完再拼回来。 | |
| 模型 | COMBO | (models/diffusion_models 下没找到 SeedVR2 权重) | SeedVR2 权重,从 models/diffusion_models 里文件名带 seedvr 的里面选。 **7B 全面强于 3B**:更清楚、身份更稳、瞎编更少。 实测(RTX 5060 Ti):单帧 1280×768 7B 6~7s / 3B 3.7s;批量跑两者几乎一样快(瓶颈在 VAE 分块,不在采样)。 所以默认 7B。3B 只在小显存或赶时间时用。 |
| 放大倍数 | FLOAT | 2.001–4 | 目标分辨率 = 原分辨率 × 这个倍数(对齐到 16 的倍数)。 **SeedVR2 自己不缩放**:先用 lanczos 把画面放到这个尺寸,模型再在这个尺寸上把细节重建回来。 2 倍性价比最高;4 倍靠的是模型「补」,不是真信息。 填了「目标短边」时这一项不生效。 ⚠️ **内存和耗时都按倍数的平方涨**(像素数 = 宽×高×倍数²)。 10 秒(240 帧)768p 竖屏按公式算:1.5 倍 ≈ 15 GB 内存 / 25 分钟;2 倍 ≈ 25 GB / 45 分钟;3 倍 ≈ 53 GB / 101 分钟。 ⚠️ 这里的「内存」是**整条链**(含下游人脸精修),超过可用内存的 90% 节点会直接拒绝开跑,不让你跑到一半卡死。 |
| 目标短边 | INT | 00–8192 | 按短边定目标尺寸。**0 = 关掉,用放大倍数**。 填 720 / 1080 / 1440 这类值,不管原图多大都拉到这个短边,长边按原比例走。 适合「统一输出到 1080p」这种活。 ⚠️ 宽高会对齐到 16 的倍数,填 1080 实际出 1088(差 0.7%)。 |
| 分块帧数 | COMBO | 自动 | 一块多少帧。必须是 **4n+1**(1/5/9/13/17/21…),因为 SeedVR2 的 VAE 时间下采样是 4 倍。 **自动(默认)**:按这一段的实际帧数,在 21~41 之间挑一个「重叠重算浪费最小」的 4n+1。 例:120 帧 → 37 帧/块 × 4 块(浪费 1.23×);240 帧 → 37 × 8;262 帧 → 41 × 8(1.25×)。比固定 21 帧快 **20% 左右**。 **选了具体数字就按那个跑**,不合规的数会往下取整到 4n+1。 **越大越快**:块间重叠的那几帧每块都要重算,块越大占比越低。 实测 2 倍 1408×2560、30 帧:21 帧/块 9.27 s/帧,41 帧/块 **7.64 s/帧**(快 18%),峰值显存只从 9.11 涨到 9.36 GiB。 显存紧就选 9 / 13;整段不超过一块时没有拼缝。 |
| 块间重叠帧 | INT | 80–64 | 相邻两块重叠多少帧,用 Hann 曲线交叉淡化接缝。 默认 8。0 = 直接硬拼,长镜头可能看到一闪。 最多取「分块帧数」的一半 —— 再多步长会塌到 1 帧,等于每帧都重算一整块,白烧算力。 重叠越大接缝越稳,但重复算的帧也越多。 |
| 色彩校正 | COMBO | lab | 把生成结果的色调拉回缩放后的原图,防止偏色。 **lab** = 在 CIELAB 空间搬颜色,最贴原片(默认); **wavelet** = 只搬低频颜色,高频细节留超分的; **adain** = 只对均值方差,最快,但可能整体染色; **none** = 不校正,只做几何对齐。 |
| 显存自适应 | BOOLEAN | true | 打开时,某一块显存不够会自动把块减半重试,并记住这个更小的大小继续跑后面的块。 关掉就严格按「分块帧数」跑,显存不够直接报错。 大分辨率(4K)或小显存卡建议开着。 |
| 随机种子 | INT | 420–2147483647 | 一步采样的初始噪声种子。**42 没有任何技术含义** ——它只是《银河系漫游指南》里的梗,社区当固定默认值用,好处是「同参数同结果」,方便 A/B 对比时排除随机因素。 这里种子对结果的影响**很小**:一步采样、且模型被输入画面的 latent 强约束,噪声只决定那一次前向的起点。 真要挑刺才去换它;正常跑就留 42。 |
| MODELopt | MODEL | 接一个已经加载好的 SeedVR2 模型,就不按「模型」那一项去读了。 一般不用接 —— 节点自己会缓存。 | |
| VAEopt | VAE | 接一个已经加载好的 SeedVR2 VAE。不接就自动找 models/vae 下文件名带 seedvr 的那个。 |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| 处理后帧 | IMAGE | — |
| 处理报告 | STRING | — |
| 放大倍数 | FLOAT | — |