SeedVR2-1.4B: 6-Layer Distilled Upscaler With ComfyUI Support
A 1.4B, 6-layer distillation of SeedVR2-7B brings one-step 4x/8x upscaling to low-VRAM machines, with a drop-in ComfyUI custom node and ready-made workflow.
SeedVR2-1.4B is a 1.44B-parameter, 6-layer knowledge distillation of ByteDance-Seed/SeedVR2-7B (the sharp EMA variant). It is a one-step diffusion image upscaler that cuts the teacher down from 8.2B parameters / 15.35 GB to 1.44B parameters / 2.69 GB, with a full ComfyUI support folder so the distilled architecture loads out of the box.
Why Distill to Six Layers
SeedVR2-7B's 36 transformer blocks are too heavy for many machines: 15.35 GB of fp16 weights and a 14–16 GB peak put it out of reach on consumer laptops, integrated GPUs, and phones. The 1.4B student keeps the same one-step diffusion formulation but uses only 6 transformer blocks, which cuts both the resident weights and the activation working set.
| SeedVR2-1.4B | SeedVR2-7B (teacher) | Advantage | |
|---|---|---|---|
| Transformer layers | 6 | 36 | 6× fewer |
| Parameters | 1,442,608,252 | 8,239,608,232 | 5.7× fewer |
| Weights on disk (fp16) | 2.69 GB | 15.35 GB | 5.7× smaller |
| 512→2048 (4×) | 20.2–22.5 s | 33.0–38.4 s | ≈1.6× faster |
| 512→2048 peak RAM | 4.6 GB | 14.2–16.4 GB | ≈3× lighter |
| 512→4096 (8×) | 54.4–55.2 s | 257–307 s | 4.7–5.6× faster |
Measured on an Apple M2 Ultra (64 GB) via MLX, model load included.
At 4× the student tracks the 36-layer teacher closely on structure and actually delivers more edge energy: outputs read crisp and detailed rather than soft. If you are memory-bound, latency-bound, or batching a lot of images, this is the model to reach for.
ComfyUI Support (Drop-In Folder)
ComfyUI supports SeedVR2 natively, but it identifies the architecture by inspecting the checkpoint and only recognizes blocks.31 (3B) and blocks.35 (7B). The 1.4B model has six blocks, so it will not load until ComfyUI is taught about it. The repo ships a drop-in support folder that fixes this without editing any ComfyUI source files:
- Copy
comfyui/ComfyUI-SeedVR2-1.4B/intoComfyUI/custom_nodes/and restart - Use the ComfyUI variant checkpoint:
seedvr2_distill_6L_1.4B_sharp_fp16_comfyui.safetensors(2.69 GB); it carries thepositive_conditioning/negative_conditioningtensors ComfyUI's loader requires - Pair it with the SeedVR2 VAE (
ema_vae_fp16.safetensors, 0.47 GB) - Drag in the ready-made workflow
seedvr2_1.4b_upscale_image.json: Load Image → Upscale Image (SeedVR2 1.4B) → Save Image, one step, prompt-free
Optional Distilled Decoder
The repo also includes an optional distilled VAE decoder (ema_vae_distill_fp16.safetensors, 0.19 GB) that is 112× smaller and ~13× faster to decode, at the cost of a mild low-pass effect. Swap it in via the ComfyUI-SeedVR2-1.4B-TinyVAE folder if you want the fastest possible decode.
Availability
- Model: lvladikov/SeedVR2-1.4B (Apache-2.0)
- Teacher: ByteDance-Seed/SeedVR2-7B
- ComfyUI reference weights: Comfy-Org/SeedVR2
- Upscaling is prompt-free: no text encoder or config download needed, just the transformer and the VAE
Comments
Sign in with GitHub to join the discussion.