Steppe Latent Upscaler
Steppe generative latent upscalers: 2x resolution and 2x frame rate in latent space for Wan 2.1 / Wan 2.2 14B / Qwen-Image latents, decoded directly by the VAE, no extra sampling pass.
ComfyUI-Steppe
Steppe is a family of generative latent upscalers: small generative models trained only for upscaling, which run directly on VAE latents. The upscaled latent is decoded straight by the VAE: no second sampling pass of the base model, no hires fix, no prompt.
- 2x resolution of images and videos, e.g. 512 → 1024 or 1024 → 2048, and chained 2x → 2x.
- 2x frame rate of videos (latent frame interpolation), also combined with 2x resolution.
- One step: about 45 ms for a 1024 px image with the default model. Most of the time goes to the VAE decode.
- Sharpness control through the high frequency feature (hff).
- Works with latents of the Wan 2.1 VAE, which is used by Wan 2.1, Wan 2.2 14B, Qwen-Image and Qwen-Image-Edit, among others.
Models: TensorForger/Steppe-Wan2.1 on Hugging Face. The nodes download them on first use.
Installation
- ComfyUI-Manager: search for "Steppe" and install.
- Comfy CLI:
comfy node install steppe - Manual:
cd ComfyUI/custom_nodes git clone https://github.com/TensorForger/ComfyUI-Steppe pip install -r ComfyUI-Steppe/requirements.txt
The nodes download models into ComfyUI/models/latent_upscale_models on first use. You can also download them
yourself from Hugging Face into that folder.
Models
| file | steps | size | use |
|---|---|---|---|
| steppe_wan_2_1_200M_1step.safetensors | 1 | 445 MB | default: the quality of the base model in one step |
| steppe_wan_2_1_200M_base.safetensors | 1-200 (20) | 444 MB | most control: steps, CFG on hff, chained upscales |
| steppe_wan_2_1_20M_1step.safetensors | 1 | 39 MB | lite: very large outputs (2048 → 4096), real time |
Nodes
| node | what it does |
|---|---|
| Load Steppe Upscaler | Loads a model and downloads released models if missing. dtype: auto (bfloat16 where supported) or float32. |
| Steppe Upscale Latent | Upscales a LATENT: spatial 2x resolution, temporal 2x frame rate, hff, guidance, steps (base model only), seed. |
| Steppe Upscale Latent (Advanced) | Adds negative_hff, shift, noise_scale and temporal windows for long videos. |
| Steppe Measure HFF | The hff value of an image, e.g. to match the sharpness of a reference. |
Basic graph: KSampler → Steppe Upscale Latent → VAE Decode. Images and videos use the same node.
Parameters
- hff (high frequency feature,
0...1, default 0.9) is the target amount of fine detail and sharpness.-1turns hff off and lets the model choose. The values are absolute: Steppe Measure HFF on a photo at the target resolution gives a value that matches it. - guidance:
- 1-step models: distilled guidance, trained on 1-3 (default 2).
1turns it off. It is a model input, so it costs nothing. - Base model: classifier free guidance between
hffandnegative_hff(default: no hff). The effect is subtle, so use 4-8 for a visible change. Guidance above 1 costs two passes per step.
- 1-step models: distilled guidance, trained on 1-3 (default 2).
- temporal: T latent frames become 1 + 2 (T - 1), i.e. 81 → 161 pixel frames for Wan. The first frame is kept.
- temporal_window (Advanced): long videos run without windows, since memory is handled by chunking. Windows of about 8-16 source latent frames keep each piece close to the training clip length.
Tips
- Chained upscaling 512 → 1024 → 2048: use two Steppe nodes. The base model with hff gives the most control over the second step. The 20M model is fast enough for the 2048 → 4096 step.
- Decoding large outputs: use
VAE Decode (Tiled)for 2048 px images and upscaled videos. - Video: 81 frames at 480x832 → 960x1664 at 2x frame rate takes 3.5 s and 9 GB peak VRAM with the 1-step model (RTX 5090), plus the VAE decode.
- Not supported: the Wan 2.2 5B (TI2V) VAE (48 channels), and FLUX, SD3 and SDXL latents. 16 channel FLUX.1 / SD3 latents pass the channel check but give wrong results.
Benchmark
1000 photos (Axis-V1), 512 → 1024 from the Wan 2.1 latent only (no prompt), patch FID on native 256 px tiles (lower is better). Latency at batch 1 on an RTX 5090, VAE decode included.
| solution | patch FID | latency | |---|---|---| | Steppe 200M base, 20 steps | 4.31 | 1051 ms | | Steppe 200M 1-step | 4.36 | 179 ms | | Steppe 20M 1-step | 5.85 | 141 ms | | Wan2.1-VAE-upscale2x decoder (spacepxl) | 6.33 | 29 ms | | SPAN 2x after decode | 11.19 | 35 ms | | Real-ESRGAN x2 after decode | 13.83 | 57 ms | | bicubic after decode | 14.13 | 30 ms | | latent upscalers made for hires fix (SesquiLSR, NN latent) | 29-51 | ~135 ms |
Python
The steppe package is pure PyTorch and does not need ComfyUI:
pip install "steppe[standalone] @ git+https://github.com/TensorForger/ComfyUI-Steppe"
from steppe import UpscaleSettings, load_upscaler
upscaler = load_upscaler("steppe_wan_2_1_200M_1step.safetensors") # downloads from Hugging Face
upscaler.model.to("cuda")
# raw (unnormalized) Wan 2.1 VAE latents, [B, 16, T, H, W] or [B, 16, H, W]
upscaled = upscaler.upscale(latents, UpscaleSettings(spatial=True, temporal=False, hff=0.9, guidance=2.0))
Related
- comfyui-flow-upscaler: the ancestor of Steppe, a FLUX.2 latent upscaler.
- LUA: latent upscale adapters: single pass regression upscalers. Steppe is generative instead.
License
MIT (see LICENSE). Steppe works on latents of the Wan 2.1 VAE (Apache-2.0, Wan Team, Alibaba Group); see NOTICE.