seams too much in comfyui
1K -4K seams too much , 512-2k is fine.
Both of those are 4×, so it's the output size rather than the factor — 512→2048 is 4.2 MP, 1024→4096 is 16.8 MP. The model itself handles large inputs fine (my best-measured line is actually 2048→8192), so this is almost certainly the VAE tiling in the workflow rather than the weights.
ComfyUI decodes the VAE in tiles (tile_size 512, overlap 128) and cross-fades them. At 4 MP that's a few tiles; at 17 MP it's several times more, so seams that were invisible start showing.
Try raising overlap from 128 to 192 or 256 — on both VAEEncodeTiled and VAEDecodeTiled inside the Upscale subgraph. That widens the cross-fade between tiles. It's a safe change: overlap affects the tile stride, not the tile size, so each tile costs the same and you just process a few more.
Don't raise tile_size to fix it. This VAE runs self-attention across the whole tile, so the attention matmul grows with the square of the tile's token count — doubling tile_size is ~16× the attention work. Depending on your backend you'll get an out-of-memory error or a hard crash, and there's no speed benefit to offset it.
If seams persist, split the job into two smaller passes — 1024→2048, then 2048→4096 — instead of one 4× jump.
thanks for your reply, but comfyui int8 model is normal using the default encode and decode tile , but yours not. I thinks something must has been broken
You're right, I have double checked, it's the model, not your settings — my tiling answer was wrong, sorry.
This is a 6-block 1.4B distillation of the 36-block 7B. SeedVR2 blends its attention windows on every other block, so the 7B has 18 blending blocks and this one has 3. The number of windows scales with output size, so at 16 MP (1024→4096) there are far more seams to blend and far less blending to do it. Below roughly 4–8 MP output it doesn't show; above that it does.
For 1024→4096, either:
run it in two passes — 1024→2048, then 2048→4096 (this works, each pass stays in the good range), or use the official 3B or 7B for single-pass jobs that large. Sorry if the answer disappoints you. Just being honest.
Where my model actually wins — 2.4× smaller, faster, lower peak memory, and competitive for the size results. I built it mostly so it can fit on low res systems and edge devices.