ComfyUI-ASASR

ComfyUI-ASASR

ComfyUI node for ASASR, a diffusion-based 4x image super-resolution model built on a FLUX.1-dev backbone with a dual-LoRA design (SR + DPO adapters) and OminiControl-style image conditioning.

• Upstream project: wafer-bob/ASASR
• Model weights: wafer-bob/ASASR on Hugging Face

All credit for the ASASR model, training, and inference code goes to its original authors (wafer-bob / Kuaishou Technology). This repository is a ComfyUI integration of their published model and weights.

Non-commercial use only. The ASASR model weights and the code derived from the upstream repository (asasr/vendor/, asasr/color_fix.py) are licensed under CC-BY-NC-4.0. They may not be used commercially. See License below.

What it does

One node, ASASR Upscale (4x), takes a FLUX.1-dev MODEL, CLIP and VAE from ComfyUI's standard loaders plus an IMAGE, and returns a 4x upscaled IMAGE (batch-aware). Because it runs on ComfyUI's own model types, it:

• works with bf16, fp8 and GGUF-quantized FLUX.1-dev checkpoints, so you can pick the precision your GPU can hold;
• reuses the FLUX.1-dev files you already have instead of downloading a separate copy;
• is loaded and offloaded by ComfyUI's own model management, like any other model in your workflow.

The input image is split into overlapping low-resolution tiles, each tile is bicubic-upscaled to a 512px condition image, denoised with FLUX + the ASASR LoRAs, and the resulting high-resolution tiles are feather-blended back together. A global color-fix pass (AdaIN or wavelet) corrects for drift against a bicubic reference of the whole image.

This is a research-grade port of an academic model. Expect it to be slow — each 128px tile takes a full 28-step diffusion pass.

Results

All examples below were produced end to end through the ComfyUI API with the default settings (28 steps, guidance 3.5, tile_size 128, tile_overlap 16, color_fix adain). Inputs are images from the Kodak Lossless True Color Image Suite, downscaled 4× with bicubic filtering, so the original image serves as ground truth.

Before / after (left: 192×128 input, shown with nearest-neighbor scaling; right: ASASR 4× output):

Compared with bicubic upscaling and the original image, with zoomed crops in the bottom row:

ASASR is a generative model: it reconstructs plausible fine detail such as window frames, lettering and textures instead of interpolating blur. Small details that are lost entirely in the input, like thin wires or exact letter shapes, are reinvented rather than recovered, so the output is not a faithful reconstruction of the original.

The bf16, fp8 and GGUF Q8 variants produce visually equivalent results; see Performance.

Requirements

• ComfyUI, running on Python >= 3.10.
• An NVIDIA GPU with CUDA. Peak VRAM is about 35 GB with a bf16 checkpoint and about 25 GB with fp8 or GGUF Q8 (see Performance). Other ComfyUI backends (Apple Silicon/MPS, CPU) have not been tested.
• FLUX.1-dev model files (see Models).
• The two ASASR LoRA adapters (~885 MB SR LoRA + ~111 MB DPO LoRA), which auto-download on first run (see LoRA weights).

Installation

Option A: ComfyUI Manager / Comfy Registry

Search for "ASASR" in ComfyUI Manager, or install it with comfy-cli:

comfy node install comfyui-asasr

Option B: Manual

cd ComfyUI/custom_nodes
git clone https://github.com/claussteinmassl/ComfyUI-ASASR.git
cd ComfyUI-ASASR
pip install -r requirements.txt

Run pip install using the same Python environment ComfyUI itself uses (e.g. activate ComfyUI's venv first, or use its embedded Python).

Restart ComfyUI afterward.

Models

ASASR runs on FLUX.1-dev, loaded through ComfyUI's standard loader nodes.

FLUX.1-dev is a gated model on Hugging Face. Accept the license at huggingface.co/black-forest-labs/FLUX.1-dev (requires a free Hugging Face account), then download from that repository with your own Hugging Face access token:

| File | Put it in | Load it with | | --- | --- | --- | | flux1-dev.safetensors | models/diffusion_models/ (or models/unet/) | UNETLoader (Load Diffusion Model) | | ae.safetensors | models/vae/ | VAELoader |

Text encoders: clip_l.safetensors and a T5-XXL file (e.g. t5xxl_fp16.safetensors or t5xxl_fp8_e4m3fn.safetensors) from comfyanonymous/flux_text_encoders, placed in models/text_encoders/ and loaded with DualCLIPLoader (type flux).

Lower VRAM: set UNETLoader's weight_dtype to fp8_e4m3fn to load the bf16 checkpoint in fp8, or use a GGUF-quantized FLUX.1-dev checkpoint with ComfyUI-GGUF's UnetLoaderGGUF (install that node pack separately). Any loader that produces a FLUX.1-dev MODEL works. Quantized conversions of FLUX.1-dev remain subject to the FLUX.1-dev license.

LoRA weights

The SR and DPO LoRA adapters are downloaded automatically on first use (via huggingface_hub), into:

<ComfyUI>/models/loras/ASASR/
  sr_lora/pytorch_lora_weights_v2.safetensors
  dpo_lora/adapter_model.safetensors

To place them manually instead (e.g. on an offline machine), download the same two files from huggingface.co/wafer-bob/ASASR and put them at those exact relative paths under <ComfyUI>/models/loras/ASASR/. Set auto_download to false on the node once the files are in place if you want to skip the network check entirely.

ASASR Upscale (4x)

Takes an image batch and returns a 4x upscaled batch (IMAGE → IMAGE).

| Input | Type | Default | Notes | | --- | --- | --- | --- | | model | MODEL | — | FLUX.1-dev UNet from UNETLoader, ComfyUI-GGUF's UnetLoaderGGUF, or any other loader producing a FLUX ModelPatcher. | | clip | CLIP | — | From DualCLIPLoader (clip_l + t5xxl, type flux). Used only to encode the empty text prompt — ASASR conditions on the LR image, not on text. | | vae | VAE | — | From VAELoader, the FLUX ae.safetensors. | | image | IMAGE | — | Batch of images to upscale. | | steps | INT | 28 (range 1–100) | Diffusion steps per tile. | | guidance | FLOAT | 3.5 (range 0.0–20.0, step 0.1) | Classifier-free guidance scale. | | seed | INT | 0 | Base seed; each tile uses a deterministic offset seed, so results are reproducible per batch and per tile layout. | | sr_lora_scale | FLOAT | 1.0 (range 0.0–2.0, step 0.05) | Scale of the internally-applied SR LoRA delta. See ASASR LoRAs are applied internally below. | | dpo_lora_scale | FLOAT | 1.0 (range 0.0–2.0, step 0.05) | Scale of the internally-applied DPO LoRA delta. Same note applies. | | condition_scale | FLOAT | 1.0 (range 0.0–4.0, step 0.05) | OminiControl condition attention scale. Values other than 1.0 allocate a dense attention bias, which is memory-expensive at production resolution; 1.0 is the fast path. | | tile_size | INT | 128 (range 64–512, step 16) | LR-side tile edge length, in pixels. See Tiling. | | tile_overlap | INT | 16 (range 0–63) | LR-side overlap between adjacent tiles. Internally clamped to at most half of tile_size. | | color_fix | combo | adain | adain, wavelet, or none. Applied globally after tiles are merged. | | auto_download | BOOLEAN | true | Auto-download the ASASR LoRA weights (see LoRA weights) if missing. |

ASASR LoRAs are applied internally, not via a LoRA loader

The SR and DPO LoRA adapters are downloaded automatically and applied internally, inside the node, as branch-gated deltas restricted to the OminiControl condition branch.

Do not load the ASASR LoRA files with a normal LoRA loader (e.g. LoraLoaderModelOnly) and feed the patched model into this node. A regular LoRA loader has no notion of which branch a delta belongs to and would apply it to the main FLUX branch as well, which deviates from the regime ASASR was trained under and will produce incorrect results. Let sr_lora_scale / dpo_lora_scale on this node control the adapter strength instead.

Memory management

Model loading and offloading is entirely ComfyUI's own (load_models_gpu) — the model behaves like any other model in a ComfyUI workflow and is subject to ComfyUI's normal VRAM management and "Free model and node cache" behavior. In addition to the loaded FLUX model, the node keeps one LoRA delta store resident on the compute device, roughly ~2.1 GB in fp32.

Known deviations from upstream

• T5 empty-prompt padding: ASASR conditions on an empty text prompt. Comfy's T5 tokenizer convention pads/truncates to 256 tokens, while the upstream reference implementation uses 512. The expected impact is negligible, since the prompt is empty either way and image conditioning dominates.
• condition_scale != 1.0 allocates a dense attention bias tensor, which is memory-expensive at production (non-tile) resolutions. Leave it at 1.0 unless you specifically need to tune this and can afford the memory cost.
• No ComfyUI attention patches / ControlNet: the forward pass calls comfy's Flux block modules directly rather than going through comfy's transformer_options hook, so attention patches, transformer_options, and ControlNet do not apply to this node.

Numerical parity

The node's forward pass is verified against the upstream diffusers reference implementation (vendored in asasr/vendor/, used only by the test suite) with tiny twin models (matching architecture, randomly initialized to keep the test fast) covering the base FLUX forward, the LoRA-patched forward, the OminiControl condition branch, and the condition_scale attention bias. Maximum observed deviation is approximately 4e-7, well under the 1e-4 parity gate used by the test suite.

Performance

Measured on an NVIDIA RTX PRO 6000 Blackwell (96 GB) with ComfyUI b5cc883, PyTorch 2.14 / CUDA 13.0, default settings (28 steps, tile_size 128). Every variant ran in a fresh ComfyUI process; VRAM is the device-wide peak reported by nvidia-smi, including the text encoders and VAE.

| Checkpoint | Time per tile | 768×512 output (2 tiles) | 1536×1024 output (12 tiles) | Peak VRAM | | --- | --- | --- | --- | --- | | FLUX.1-dev bf16 | 5.5 s | 11.3 s | 65.4 s | 35.3 GB | | FLUX.1-dev fp8 (e4m3fn) | 7.0 s | 14.3 s | 83.6 s | 24.6 GB | | FLUX.1-dev GGUF Q8_0 | 8.5 s | 17.4 s | 101.9 s | 25.1 GB |

Times are for a warm model; the first run after starting ComfyUI adds roughly 10–20 s of model loading. Runtime scales linearly with the number of tiles.

• Quality: all three variants produce visually equivalent output. The bf16 variant matches the upstream diffusers reference implementation up to bf16 rounding.
• fp8 saves memory, not time: on this GPU the fp8 weights are cast per layer during the forward pass, so fp8 is slower than bf16.
• Smaller GPUs: with ComfyUI limited to a 24 GB budget (--reserve-vram 72 on the 96 GB card) the fp8 and GGUF variants ran at unchanged speed, and fp8 also stayed within a 16 GB budget. These were simulated limits on a large card, not tests on real 16/24 GB GPUs.

Tiling

ASASR was trained on 128px → 512px tiles, so tile_size = 128 (the default) is the model's native, most faithful operating point.

The input image is padded and split into a grid of overlapping tiles of tile_size x tile_size pixels (LR side). Each tile is bicubic-upscaled to 512x512 to form the FLUX conditioning image, denoised for steps steps, and the resulting 4x tile is blended into the output using a feathered mask over the overlap region (tile_overlap) so seams don't show. After all tiles are merged, a global color-fix pass (AdaIN or wavelet, from the StableSR color-fix method) aligns the result's color statistics with a bicubic upscale of the original input.

When to change the defaults:

• Larger tile_size processes fewer tiles per image (faster overall), but moves away from the model's trained regime and can lose fine detail fidelity. Generation always runs at the model's native 512px condition resolution: for tile_size above 128, the per-tile result is bicubic-resized from that 512px render, so tile sizes above 128 do not add model-generated detail beyond a 512px-per-tile render. tile_size = 128 is the faithful, native setting.
• Larger tile_overlap reduces visible seams at the cost of more redundant diffusion work (more overlapping pixels are computed twice). It cannot exceed half of tile_size.
• color_fix: none skips the global color correction if you plan to do your own color grading downstream, or if you notice color-fix artifacts on unusual inputs.

A 512x512 input at the default tile_size=128 already produces 16+ tiles (before overlap padding), each requiring a full diffusion pass — this is the main cost driver, not image resolution alone.

Example workflow

A minimal graph (LoadImage + UNETLoader + DualCLIPLoader + VAELoader → ASASR Upscale (4x) → SaveImage) is included at example_workflows/asasr_upscale_example.json:

Load it in ComfyUI via Workflow → Open (or drag the file onto the canvas). Once the node pack is installed, it also appears in ComfyUI's template browser under ComfyUI-ASASR. Swap the UNETLoader node for ComfyUI-GGUF's UnetLoaderGGUF to use a GGUF-quantized checkpoint.

License

• ComfyUI integration code in this repository (everything outside asasr/vendor/ and asasr/color_fix.py): MIT. See LICENSE.
• asasr/vendor/ and asasr/color_fix.py, derived from the upstream ASASR repository, and the ASASR model weights (downloaded separately from Hugging Face): CC-BY-NC-4.0 — non-commercial use only. See asasr/vendor/LICENSE. The model weights additionally inherit the non-commercial terms of the underlying FLUX.1-dev model.

Because the model weights and part of the code are non-commercial, the combined package (this repository plus the downloaded weights) is for non-commercial use only, regardless of the MIT license on the node code itself.

Credits

• ASASR model, LoRA weights, and inference code: the ASASR authors (wafer-bob) and Kuaishou Technology — github.com/wafer-bob/ASASR, huggingface.co/wafer-bob/ASASR.
• Color-fix routines (AdaIN / wavelet reconstruction): adapted from StableSR's color-fix implementation by Li Yi — pkuliyi2015/sd-webui-stablesr.
• Example images: Kodak Lossless True Color Image Suite (kodim03, kodim08, kodim23), released by Eastman Kodak for unrestricted use.
• FLUX.1-dev: Black Forest Labs — huggingface.co/black-forest-labs/FLUX.1-dev.