SeedVR2 Block Swap Memory Fix | Low VRAM Upscale Checklist
Fix SeedVR2 low-VRAM crashes with block swap, tiling, offload, smaller batches, and a browser fallback when ComfyUI tuning blocks delivery.
Door 1
Do not want to install? Use the browser
Skip ComfyUI when you only need one finished asset, when model downloads fail, when a custom node is missing on a deadline, or when an 8GB card cannot hold a useful batch. Local control is for repeatable graphs; one delivery belongs in the browser.
Door 2
Keep installing locally
Continue the checklist, model folders, and VRAM notes on this page.
First-party before/after SeedVR2 proof
This comparison uses a first-party demo source and output published on this page. It is not a third-party customer asset or a promise about every source type.
- Block swap helps low-VRAM machines fit larger workflows, but it adds setup time and slower generation.
- Reduce batch size, output size, and tile size before changing many ComfyUI dependencies at once.
- Use online image or video lanes when you need a result faster than a local dependency or memory fix.
Direct answer
BlockSwap helps SeedVR2 fit into low VRAM by moving model blocks through system memory, but it is not a speed or quality upgrade. Turn it on only after you confirm the failure is a CUDA OOM or model-fit issue. If the page is about one asset instead of a repeatable local workflow, online SeedVR2 is the cleaner fallback because it avoids the swap-tuning loop.
- Use BlockSwap for repeatable OOM, not for missing nodes or model path errors.
- Expect slower runs as more blocks are swapped through CPU memory.
- Keep VAE tiling separate because it solves encode/decode memory pressure.
Confirm the crash is actually a memory problem
BlockSwap should be a targeted fix, not the first setting you touch. Read the terminal log and confirm the failure is an out-of-memory error during model inference. A red node, missing Python package, wrong model directory, unsupported attention backend, or broken workflow JSON will not be fixed by swapping blocks. Solve the first real error in the log before changing memory settings.
If the job fails only at a larger target resolution, reduce output size and batch first. If the same source runs at 720p or 1080p but fails at 4K, memory pressure is likely. If no resolution runs, the setup may be broken earlier in the chain and should be repaired before BlockSwap is considered.
- Look for CUDA OOM or memory allocation errors.
- Fix missing nodes before tuning memory.
- Run a smaller resolution to confirm the workflow itself works.
- Save the first working settings as a rollback point.
Practical BlockSwap settings for low VRAM
BlockSwap works by keeping only part of the transformer workload in VRAM and moving the rest through CPU memory or another device. Start with CPU offload and a moderate number of swapped blocks. Increase the number only when the same job still fails. More swapped blocks can reduce peak VRAM, but every extra transfer costs time and can make the workflow feel stalled on slower systems.
The NumZ community SeedVR2 node exposes BlockSwap on its DiT loader, while VAE tiling is configured separately on the VAE loader. Keep those two ideas separate: BlockSwap helps the transformer model fit, and VAE tiling helps encode or decode large frames. Neither setting turns the workflow into official 8GB support.
- Set offload to CPU before enabling BlockSwap.
- Increase swapped blocks only after a repeatable OOM.
- Use VAE tiling for encode/decode OOM, not model-load OOM.
- Expect slower output when swapping is high.
Quality and speed tradeoffs
BlockSwap does not automatically lower visual quality, but the settings people combine with it often can. Overly small tiles, too little overlap, very low batch size, and aggressive output jumps can create seams, flicker, or invented texture. When you change a memory setting, keep the model, source, seed, and target resolution stable so the comparison is meaningful.
Use online SeedVR2 when the business problem is delivery, not research. A local BlockSwap workflow is worth documenting if you will run it repeatedly. For one-off client previews, social clips, or a product image that needs to ship today, the browser workflow can be the better conversion path because it avoids dependency time and returns failed-job credits automatically.
Who this guide is for
Built for search visitors who need a finished upscale, local ComfyUI users choosing between native SeedVR2 and community extensions, and teams deciding whether online delivery is safer than local GPU maintenance.
Test method and evidence scope
- Published
- 2026-07-31
- Updated
- 2026-08-23
- Owned online proof
- WaveSpeed SeedVR2 image endpoint, owned SVG source, 768px input to 2048px WebP output, submitted 2026-07-20.
- Local ComfyUI evidence scope
- Use this page as the capture checklist for RTX 8GB, 12GB, 16GB, and 24GB+ local validation before publishing hardware-specific claims.
- Success standard
- 100 percent crop or timestamp review must preserve identity, text, edges, color, and motion without hiding setup failures.
GPU and parameter table
| VRAM | Model | VAE tiling | BlockSwap | Batch | Output limit | Risk |
|---|---|---|---|---|---|---|
| 8GB | 3B or GGUF Q4_K_M/Q8_0 first | On when encode/decode memory fails; start near 1024 tile and 128 overlap | Use only after a repeatable CUDA OOM or model-fit failure | Image: 1. Video: 5 when motion consistency matters | 720p-1080p tests, then 2K stills; avoid long 4K video first | Slow transfers, seams from too little overlap, repeated OOM on long clips |
| 12GB | 3B FP8/GGUF or conservative 7B test | On for 4K stills or large frames; off for small clean tests | Light swap only when peak memory proves it is needed | Image: 1. Video: 5-9 after a short test | 2K-4K stills, short 1080p video segments | Borderline 7B jobs can pass once and fail on larger aspect ratios |
| 16GB | 3B/7B FP8, then FP16 if the source deserves it | Use for large frames and 4K; keep overlap visible in notes | Usually off for image tests; use for heavier video batches | Image: 1. Video: 9-13 when stable | 4K stills and moderate video clips | Overconfident 4K/8K jumps can amplify source artifacts |
| 24GB+ | 7B FP16 or sharp variant after a 3B baseline | Optional for smaller jobs; useful for high-resolution output | Usually off; keep it as a fallback, not a default | Image: 1. Video: 13-21+ after memory logging | 4K stills, longer 1080p/4K segments, production comparisons | Bigger models can invent sharper artifacts when the source is poor |
Troubleshooting errors people actually search
SeedVR2 missing node
The workflow references a node type that this ComfyUI build cannot load. A native template usually means ComfyUI is outdated; a v2.5 community workflow may require its matching custom node.
Update ComfyUI to 0.28.0 or newer and reopen the official SeedVR2 template first. Only install the matching community node when the workflow explicitly depends on it, then restart and resolve the first terminal error.
CUDA out of memory
The graph, model precision, frame batch, or target resolution is larger than the available VRAM for the current run.
Reduce target size or batch first, then enable VAE tiling, CPU offload, GGUF/FP8 models, or BlockSwap in that order.
model not found
The node loaded correctly, but the workflow-selected model filename does not match the local folder scanned by the node.
Place DiT and VAE files in the documented SeedVR2 model folder, keep filenames unchanged, restart ComfyUI, and select the exact model in the loader.
VAE decode crash
Encoding or decoding large frames can fail separately from DiT model inference, especially at 4K or with long clips.
Enable VAE tiling, start with a conservative tile size and overlap, then increase output size only after a smaller pass succeeds.
ComfyUI Manager install failed
The node package may be blocked by Python environment mismatch, dependency conflicts, or a stale Manager cache.
Confirm the base ComfyUI workflow runs, install one node package at a time, restart, and use the browser workspace when the asset must ship before dependency repair.
Workflow decision table
| Scenario | Recommended path | Local setting | Online fallback | Risk |
|---|---|---|---|---|
| One clean still image, product shot, or AI artwork | Online image upscaler | Batch 1, 2K first, inspect 100 percent crop | /tools/upscaler | 4K can amplify artifacts when the 2K crop already fails |
| Repeatable local ComfyUI workflow | Local SeedVR2 v2.5 graph | Record DiT, VAE, precision, tiling, offload, BlockSwap, and batch | /blog/tutorials/seedvr2-v25-installation-guide-2026 | Missing nodes and model paths can consume more time than the asset is worth |
| Motion clip, reel, or compressed video | Online video upscaler | Use 4n+1 batches only if running a video graph locally | /tools/video-upscaler | Frame-by-frame image upscaling can cause temporal inconsistency |
| Old portrait, family scan, or identity-sensitive face | Photo enhancement first | Conservative face review before enlargement | /tools/photo-enhance | Sharper faces can still become the wrong identity |
Choosing between the browser and a local install
Use the online image workspace when the source is a still image, product asset, portrait, transparent PNG, or AI artwork that needs a clean 2K or 4K result without dependency repair.
Sources and update notes
Checked against the linked ByteDance, Hugging Face, Comfy-Org, and ComfyUI sources plus the displayed first-party comparison.
- ByteDance Seed official repository
Upstream source for the Apache-2.0 license, official 3B/7B checkpoints, snapshot_download examples, H100 reference environment, and documented model limitations.
- Comfy-Org ComfyUI repack
ComfyUI-formatted model files and their exact diffusion_models and VAE directories; this is a repack, not the ByteDance model publisher.
- NumZ community custom node
Community-maintained custom node used for v2.5 layout, GGUF, tiling, BlockSwap, and batch behavior; it is not an official ByteDance release.
- Official ComfyUI video upscaling documentation
Current ComfyUI documentation exposes both open-source and hosted SeedVR2 workflows and recommends testing sample clips before full processing.
- Official ComfyUI SeedVR2 workflow template
The maintained template identifies ComfyUI core 0.28.0, the 3B INT8 diffusion model, the shared VAE, exact model folders, and native SeedVR2 support.
- SeedVR2 paper
Used for the one-step video restoration framing and source-quality expectations.
- First-party SeedVR2 image comparison
Before/after image proof generated from a self-authored SVG source, with request metadata published alongside the asset.
Related SeedVR2 workflows and fixes
Try one image online
Upload one photo in the browser and see credits before you run.
SeedVR2 ComfyUI Install Guide: 3B/7B & Missing Nodes
SeedVR2 ComfyUI install: update to 0.28.0, place seedvr2_3b_int8_convrot.safetensors and the VAE in the right folders, then fix OOM and missing nodes.
Can SeedVR2 Run on 8GB VRAM? OOM Fixes That Work
SeedVR2 on 8GB VRAM in ComfyUI is a community experiment, not an official promise. Reduce resolution, try INT8 or GGUF, VAE tiling, then BlockSwap.
SeedVR2 Z-Image 4K Upscale on 8GB VRAM (Guide)
Z-Image 4K upscaling with SeedVR2 on 8GB VRAM: generate with Z-Image first, test a 2K crop before 4K, enable VAE tiling, and upscale only sources with detail.
SeedVR2 video upscaler
Use the online video upscaler and online video enhancer for duration pricing, model choice, and temporal cleanup.
Old photo enhancer
Start here for damaged portraits, scanned family photos, blur, grain, or identity-sensitive faces.