━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🔼 SeedVR2 Batch Upscaler — Sleep On It, Wake Up 4K

Drop a folder, come back to 4K

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

ComfyUI batch image upscaler powered by SeedVR2, the SOTA diffusion-based restoration model from ByteDance. Load an entire folder of QC-approved images, set the batch number, click Run once, and watch them upscale to 4K in a single session. No per-image re-queuing, no repetition—sequential images feed through the same upscaler instance to your output folder, all with automatic date-stamped organization. Built for production stock photography, concept art detail recovery, and QC'd generation batches.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

✨ Features

✅ Batch Folder Iterator — Place 10–100 images in input folder, set the batch number to match, all process sequentially without re-loading model

✅ Whole-Folder Workflow — ImageIterator node handles file discovery, sorting, and index auto-increment

✅ SeedVR2 7B fp16 (default) — Diffusion-based upscaler, SOTA quality; fp8 and 3B variants available for speed/VRAM trade-off

✅ 4K Short-Edge Target — short edge upscaled to 4096, long edge capped at 4096 (1024×1024 → 4096×4096; 1280×1600 → 3276×4096). The effective upscale factor depends on the input's aspect ratio.

✅ Block-Swap Ready — Pre-configured with blocks_to_swap=36 + CPU offload + SDPA attention for 16GB VRAM fit (requires ~33GB system RAM)

✅ Automated Output Organization — Results saved to output/ folder with automatic date stamp [YYYY-MM-DD]

✅ Torch.compile Optional — Node included but disabled by default; enable if you have triton-windows for inference speedup

✅ Metadata-Safe Export — PNG output; embedded workflow metadata has no machine paths or language-specific notes

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📦 Required Models (3 files, ~7-16 GB depending on variant)

Primary (default — recommended)

• seedvr2_ema_7b_fp16.safetensors — Main diffusion upscaler (highest quality, fits 16GB with block-swap)

• ema_vae_fp16.safetensors — VAE codec for image reconstruction

Alternative Variants (choose one — do NOT mix in same batch)

• seedvr2_ema_7b_fp8_e4m3fn.safetensors + ema_vae_fp16.safetensors — Faster, near-identical quality, lower VRAM

• seedvr2_ema_3b_fp16.safetensors + ema_vae_fp16.safetensors — Fastest option, fits 16GB without block-swap

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⬇️ Download Links (verified HuggingFace repositories)

7B fp16 (default — recommended for quality)

• seedvr2_ema_7b_fp16.safetensors — https://huggingface.co/numz/SeedVR2_comfyUI/blob/main/seedvr2_ema_7b_fp16.safetensors

• ema_vae_fp16.safetensors — https://huggingface.co/numz/SeedVR2_comfyUI/blob/main/ema_vae_fp16.safetensors

7B fp8 (faster, lower VRAM)

• seedvr2_ema_7b_fp8_e4m3fn.safetensors — https://huggingface.co/numz/SeedVR2_comfyUI/blob/main/seedvr2_ema_7b_fp8_e4m3fn.safetensors

• ema_vae_fp16.safetensors — https://huggingface.co/numz/SeedVR2_comfyUI/blob/main/ema_vae_fp16.safetensors

3B (fastest, native 16GB fit)

• seedvr2_ema_3b_fp16.safetensors — https://huggingface.co/numz/SeedVR2_comfyUI/blob/main/seedvr2_ema_3b_fp16.safetensors

• ema_vae_fp16.safetensors — https://huggingface.co/numz/SeedVR2_comfyUI/blob/main/ema_vae_fp16.safetensors

Installation Path:

• Diffusion models: ComfyUI/models/diffusion_models/

• VAE files: ComfyUI/models/vae/

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🧩 Required Custom Nodes

⚠️ Manual Install Step (do this first — it's the only one not in ComfyUI Manager):

Image_Anything (Batch Folder Iterator) — NOT in ComfyUI Manager registry

1. Open terminal/PowerShell in your ComfyUI root directory

2. Navigate to custom_nodes: cd custom_nodes

3. Clone the repo: git clone https://github.com/ComfyUI-Kelin/ComfyUI_Image_Anything.git

4. Restart ComfyUI

Then install these via ComfyUI Manager (search → install):

• ComfyUI-SeedVR2_VideoUpscaler (numz) — canonical SeedVR2 node pack, in Manager registry, searchable by name

- Install via Manager: Manager → Install Custom Nodes → search "SeedVR2" → select numz's ComfyUI-SeedVR2_VideoUpscaler → Install

• was-node-suite-comfyui (ltdrdata) — Image Save node (PNG/WebP export); MIT license, in Manager registry

- Install via Manager: search "was-node-suite" → install

Verify Installation: After restart, load this workflow in ComfyUI. If nodes resolve (no red outlines), you're good to go.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🚀 How to Use

1. Prepare Images — Place your QC-approved images (PNG/JPG, any resolution 256–4096px) in ComfyUI's ./input folder

2. Load This Workflow — Open SeedVR2_Batch_Upscale_v1.1.json in ComfyUI

3. Run It — Set the batch number next to Run to the number of images you want, then click Run once (older ComfyUI builds label this button Queue)

- Each execution advances the ImageIterator by one file, so a batch of 10 walks 10 images. Clicking Run repeatedly does not stack on current frontends — it ignores presses until the current run finishes, so use the batch number.

4. Monitor Progress — Watch the ImageIterator node report "current_index / total_count" as it walks the folder

5. Collect Output — Check ComfyUI's ./output folder for upscaled images, organized by date [YYYY-MM-DD]

Example Workflow:

- Input folder: ComfyUI/input/ (contains 5 PNG files)

- Set the batch number to 5 and click Run once

- Wait ~1 minute per image on the default 7B fp16 config (measured — see the benchmark table below); slower if your system RAM is tight

- Output appears in ComfyUI/output/[YYYY-MM-DD]/, named from the Image Save prefix (default Up VR2 [time(...)] plus a counter — edit the prefix widget on the Image Save node to change it)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⚙️ Settings & Parameters

Model Loading

• Model Path: Automatically detects from ComfyUI/models/diffusion_models/ (VAE loads from ComfyUI/models/vae/)

• Device: CUDA (GPU); falls back to CPU if CUDA unavailable

• Offload Device: CPU (system RAM) — keeps VRAM lean

Upscale Output

• Resolution: 4096 — target for the SHORT edge; aspect ratio is preserved

• Max Resolution: 4096 — cap on the LONG edge; if an image exceeds it after the short-edge target is applied, both dimensions scale down proportionally. This keeps VRAM predictable on tall or wide images.

• Batch Size: 1 per queue (sequential processing)

• Color Correction: "lab" (perceptually-aware color preservation)

• Attention Mode: "sdpa" (scaled-dot-product attention — memory-efficient)

Block-Swap (VRAM Management)

• blocks_to_swap: 36 (default, pre-configured for 16GB VRAM)

- Swaps 36 UNet blocks to system RAM to fit model on GPU

- Requires ~33GB system RAM; 64GB recommended for smooth performance

- If you have less RAM, switch to 7B fp8 or 3B variant

Torch.compile (Optional Speedup)

• Status: DISABLED by default (node present, not connected)

• Reason: Requires triton-windows, which most Windows users lack → would error on first run

• How to Enable (if you have triton): Reconnect the SeedVR2TorchCompileSettings node output → the torch_compile_args input on SeedVR2LoadDiTModel. The same input also exists on SeedVR2LoadVAEModel if you want to compile the VAE too. Can add ~20–30% speed boost, but only if dependencies are met.

Seed & Noise

• Seed: 42, with control_after_generate set to "fixed" — every run is reproducible. Set the control to "randomize" if you want variation between runs.

• Input Noise Scale: 0.05 · Latent Noise Scale: 0.0 — low values preserve detail and prevent hallucination

VAE Tiling (if image >2048px)

• Encode Tiled: Enabled (1024px tiles, 128px overlap)

• Decode Tiled: Enabled (same tiling)

→ Prevents OOM on large inputs

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

💡 Performance Tips

VRAM & System RAM Honesty

• 7B fp16 + swap36 (default, shipped) — ~16GB VRAM, ~33GB offloaded to CPU. Highest quality. Requires 64GB system RAM for smooth batch processing; 32GB will thrash and be very slow. Measured: ~62s/image (steady-state, 17-image production batch, 1280×1600→3276×4096)*

• 7B fp8_e4m3fn — ~9GB VRAM, ~18GB offloaded to CPU. Faster inference (~30% speedup estimated, not yet benchmarked). Quality near-identical to fp16. Fits 32GB system RAM comfortably.

• 3B fp16 — ~16GB VRAM, no offload needed. Fastest option (not yet benchmarked). No block-swap overhead. Fits 16GB natively.

*Measured on RTX 5080, 7B fp16 variant only, back-to-back queue (model already loaded). fp8/3B timings not yet benchmarked — estimates only.

Practical Guidance

• For highest quality + 64GB system RAM: Use 7B fp16 (default config)

• For speed + 32GB RAM: Switch model to 7B fp8 (same setup, different checkpoint file)

• For minimal VRAM/RAM: Use 3B variant (nearly as good, no block-swap delays)

• Batch 10+ images together to amortize the one-time model load

• Avoid running alongside heavy generation (e.g., Krea2 gen) — too much total VRAM pressure

• Input image quality: Sharp, well-lit images upscale better than low-contrast originals (expected for all upscalers)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🔗 My Workflow Suite

I maintain a growing library of ComfyUI workflows. Check them out:

• LTX-2.3 Image-to-Video — Lock-camera i2v with auto motion prompt (QwenVL)

• Krea2 Turbo Dual-Mode — Fast diffusion gen (text or image-to-image)

• Z-Image-Turbo — Another fast gen option with auto-prompt: https://civarchive.com/models/2742740/z-image-turbo-qwenvl-dual-mode-auto-prompt

• SeedVR2 Batch Upscaler — This workflow (batch folder upscale to 4K)

Find LTX-2.3 and Krea2 Turbo on my Civitai profile page. More coming soon — follow for updates!

GitHub Mirror: https://github.com/Thinni63/comfyui-workflows/tree/main/seedvr2-batch-upscale

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🔄 Changelog

v1.1 (2026-08-17)

• Fixed: SeedVR2 was never actually connected. The Preview node carried a phantom output slot that ComfyUI drops when loading the file, which cut the link into SeedVR2, while a stale link ID made the Save node read the preview image instead. Net effect: saved files came out identical to the input. Thanks to daze45230 for the detailed report.

• Changed: SeedVR2 is no longer hidden inside a subgraph. The model loaders, upscaler, and torch.compile node now sit directly on the canvas, so every setting is visible and editable without opening a container — and a subgraph-capable ComfyUI build is no longer required.

• Improved: Node layout reorganized into labelled groups (start guide / run settings / model setup / pipeline).

• Settings and model requirements are unchanged — drop-in replacement for v1.

v1 (2026-07-03)

• Initial release.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📝 Notes & AI Disclosure

• AI-Generated Example Images — All provided examples are AI-generated via Krea2 + upscaled with this workflow

• Hardware Tested — Verified on RTX 5080 16GB VRAM, Windows 11, CUDA 12.1+

• ComfyUI Version — No subgraph support required as of v1.1 (the graph is flat). Any reasonably recent ComfyUI with the three custom node packs installed will load it.

• Model Weights License — SeedVR2 weights are NOT distributed with the workflow; you download them separately from HuggingFace (see Download Links above)

• Metadata Safety — This workflow JSON has no machine paths or language-specific notes; it's safe to share and distribute

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⭐ Found this useful?

• Like if it saved you time

• Comment your results — I read every one

• Follow for new ComfyUI workflows, all tested on 16 GB VRAM

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Roadmap Teaser

A video-batch variant of this upscaler is coming soon. Stay tuned!

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⚖️ Model Attribution & Licensing

SeedVR2 (ByteDance Seed Team)

• License: Apache License 2.0 — https://huggingface.co/ByteDance-Seed/SeedVR2-7B

• Attribution Required: "SeedVR2 by ByteDance Seed Team, licensed under Apache License 2.0."

• Safetensors Conversion: Hosted at https://huggingface.co/numz/SeedVR2_comfyUI for ComfyUI-compatible format

• Free for commercial and non-commercial use (see license for full terms)

Custom Nodes

• ComfyUI-SeedVR2_VideoUpscaler (numz) — Check repository for license

• was-node-suite-comfyui (ltdrdata) — MIT License

• ComfyUI_Image_Anything (ComfyUI-Kelin) — MIT License

This Workflow (JSON Configuration)

• Original work by TP_AI_63 (Civitai) / Thinni63 (GitHub)

• Shared under MIT License; credit appreciated

• Model weights must be downloaded separately (not included)

All example outputs are AI-generated via Krea2 generation and SeedVR2 upscaling.