SeedVR2 7B Guide: FP8 Weights, Placement, When to Use It

SeedVR2 ships in two published sizes, and the larger one causes most of the confusion. People download the 7B checkpoint, drop it where the 3B file used to be, wait longer, and ask why the output looks the same. Sometimes it does, and that is the useful answer. This guide covers what 7B actually changes, when it is worth the extra time and memory, and what to check when the bigger model leaves the result soft.

What the 7B checkpoint is

SeedVR2 is ByteDance's diffusion restoration model, and 3B and 7B are the same design at two scales. 7B has more parameters, so the file is larger, more of it sits in memory during a run, and every sampling step costs more compute. No separate feature set, no different workflow, no extra node.

The variant discussed on this site for local work is 7B FP8. The 3B local path is the INT8 file seedvr2_3b_int8_convrot.safetensors, and it remains the low-VRAM option. Both checkpoints use the same VAE file, so switching does not mean a second download of the decoder.

7B is not a different tool. It is the same tool with a longer warm-up, a bigger memory bill, and a better chance of resolving fine texture on material that still contains it.

3B vs 7B: what you actually pay for

The comparison is a trade, not a ranking.

The gain is not spread evenly across the frame. 7B's advantage appears where the source still holds detail the 3B pass flattens: a cheek, a knit sweater, a shop sign in the background. On a 720p clip that has already been through a social platform's encoder, both checkpoints are inventing, and the bigger one invents more confidently, not more truthfully. "7B is sharper" is fair on a good source and misleading on a wrecked one.

FP8, INT8, and the "7B INT8" search

FP8 and INT8 both store weights at lower precision than FP16. The label describes what is inside the file, not a quality tier, which is why people expect INT8 to beat FP8.

On this site, the 3B local checkpoint is INT8 and the 7B checkpoint is FP8. If you arrived searching for "SeedVR2 7B INT8", you are usually asking a different question: how do I get 7B to fit in the memory I actually have? FP8 is the first step down from full precision for 7B. GGUF builds go further in the same direction, trading more speed for a smaller resident footprint; the GGUF setup guide covers that trade and the loader side in detail.

A lower-precision 7B file is not automatically better than a full-precision 3B one; precision loss and parameter count pull in opposite directions, and only a comparison on your own footage settles it.

Where the files go in ComfyUI

There are two placements; only one is usually the problem.

The 7B checkpoint is a diffusion model file too, so it belongs in the same folder as the 3B file. The model download page lists the files by variant. The VAE does not change when you switch checkpoints, which is why it gets forgotten. Keep seedvr2_ema_vae_fp16.safetensors in ComfyUI/models/vae and leave it there.

The install guide targets ComfyUI 0.28.0. A workflow written for one build can load a checkpoint and still misbehave on another, and the error points at the model rather than the version.

The other trap is stale references. Dropping the 7B file into the folder does not change what the workflow loads. The loader dropdown is the only thing that decides which file runs; a graph still pointing at the 3B filename keeps producing 3B output. Before concluding 7B made no difference, confirm the node is reading it.

Running 7B without a local GPU

The checkpoint size is only a problem if the hardware is yours. SeedVR2.net runs ByteDance's SeedVR2 models server-side, so the 7B memory bill and the ComfyUI version stop being your problem. Upload a file to the image upscaler or video upscaler, pick a target, and the job runs on managed hardware.

That path does not let you choose a checkpoint, tune block swap, or inspect the sampler graph. What it removes is the whole class of failures this article is about: wrong folder, stale loader, misplaced VAE, mismatched build.

What 7B costs, hosted and on your desk

Locally, the marginal cost of a 7B run is electricity and time, and the fixed cost is the GPU that can hold it. Hosted, the price is quoted before the job starts and does not change with the checkpoint size. Image rates are flat by target resolution:

Video bills per second of source footage, with a 5-second minimum and a 10-minute maximum per clip. On the standard model the rates are 9 credits per second at 720p and 1080p, 17 at 2K and 33 at 4K. Other models cost more; the Ultimate model is 131 credits per second at 4K. A 5-second 720p clip starts at 45 credits, and a 30-second 1080p clip is 270.

Credits come from packs or subscriptions:

Yearly subscriptions are $118.80 for Starter (9,600 credits), $238.80 for Creator (28,000) and $598.80 for Studio (80,000). A first 2K image upscale is free, and so is one standard 720p video job up to 5 seconds with a verified email, enough to test the model on your own footage.

When 7B still looks soft

If the bigger checkpoint loaded and the output is still soft, work through these before downloading anything else.

• You are comparing at fit-to-window. A full frame scaled to your monitor hides exactly the detail 7B is supposed to add. Open both results at 100% on the same crop.
• The detail was never in the source. Diffusion restoration reconstructs plausible texture; it cannot recover information the encoder discarded. On a compressed 480p file, 7B produces cleaner guesses, not real eyelashes.
• You changed two variables. 3B at 4K against 7B at 2K is not a model comparison. Match the target resolution, then change one thing.
• The graph is still on 3B. Confirm the loader dropdown, not just the file in the folder.
• The softness is motion blur. Blur that moves across frames is a capture problem; a larger checkpoint cannot undo a shutter that was too slow.
• The clip is long. Memory runs out before quality does. If a 7B run dies partway through, shorten the clip and test again.

The other thing worth checking is identity. A stronger model with a free hand on detail can change a face in ways a human notices instantly even when every edge is sharper. For a portrait or a family video, compare the face at 100% before you batch the rest.

When 3B is the better answer

• You are installing for the first time. Get a 3B run finishing end to end before adding a second checkpoint to the folder.
• You have 8GB of VRAM. The 3B INT8 path runs there with block swap; long clips still run out of memory first.
• You are processing a batch. Per-step time multiplies across every clip. If 3B already passes your quality bar, 7B is a tax with no visible return.
• The source is already clean. A sharp, well-lit, high-bitrate clip has little missing detail worth reconstructing.

7B is a detail upgrade for sources that still have detail to find, not a fix for a bad source.

Limitation

This post deliberately contains no benchmark numbers, no timing table and no VRAM figures for 7B. Exact headroom varies by ComfyUI build, node pack version, clip length, output resolution and batch size, so any single number here would be wrong on someone else's machine. The 8GB figure applies to the 3B INT8 path with block swap, and long clips still run out of memory first. Whether 7B is worth it on your footage cannot be settled from this page: run one crop through both checkpoints at the same target resolution, open the results at 100%, and compare a face and a texture. If the difference is not visible there, it will not be visible in the batch.