SDXL ChordEdit (MODEL/CLIP/VAE)

ComfySDXLChordEdit (displayed as "SDXL ChordEdit (MODEL/CLIP/VAE)") is a text-driven image editor that replaces the usual img2img guessing game with one idea: whatever direction separates two prompts in the model's conditioning space is the direction your image should move. Feed it a photo, tell it the source prompt is "a photo of a person" and the target is "a watercolor painting of a person", and it walks the image's latent toward that change. No KSampler, no denoise slider, no separate Conditioning Set Area nodes - one node takes MODEL, CLIP, VAE, an image, and two sentences, and hands you the edited image plus the latent.

It's from the Fengxiaoxiao-001/ComfyUI-FastTool pack ("XiaoXiao" is the author's handle, and it's in the XiaoXiao/ChordEdit category). This is the research-y corner of an otherwise VRAM-focused toolkit, and it's honestly a niche tool. "ChordEdit" has essentially no community footprint on Reddit - this is a small author's take on the chord/direction-editing family of methods, not a household name like InstructPix2Pix. If you're comfortable tuning a KSampler you may not need it. But it is genuinely one-node-simple for prompt-consistent edits, which is rare.

How it actually works

The node encodes your image into SDXL latent space with proper crop conditioning, encodes both prompts, and computes the edit direction: the difference between the target and source conditioning (both cross-attention and pooled output). That's the chord - two prompt points, one direction.

The direction is noisy, so it estimates it the Monte Carlo way. It creates noise_samples random noises (seeded), and at each step from t_start (default 0.85) down to t_end (default 0.15) it runs the UNet on each noise at two nearby noise levels (t and t - t_delta), each under both the source and target prompts. That's four model inputs per noise per step - which is exactly what the unet_batch_size tooltip means by "each noise generates four sets of model inputs". The averaged difference gets scaled by direction_gain and step_scale, then added to the latent, step by step.

The inputs that matter

You will actually set most of these, but a few carry the tuning weight:

• source_prompt / target_prompt - the two ends of the edit. Keep them structurally similar ("a photo of a person" → "a watercolor painting of a person"); if they're identical the node raises an error because there's no direction to walk.
• noise_samples (default 4) - more samples = smoother, less noisy direction, but 4 UNet passes per sample per step. This is where the VRAM and time go.
• direction_gain (default 2.0) - the tooltip is the author's own advice: 1.0 is the original formula, but SDXL works better from 2.0. Turn it up if the edit is too weak, down if it overshoots.
• step_scale (default 0.35) - how far each step pushes along the direction.
• n_steps, t_start, t_end, t_delta - the sampling window. The defaults (8 steps, 0.85→0.15) are a sane starting point.
• width / height (default 1024×1024, SDXL native) with center_crop and force_square_when_cropping on by default - the latter forces a square target to match the original ChordEdit preprocessing, so leave it unless you have a reason not to.

Outputs are image (the edited result, decoded) and latent - wire the latent into a KSampler if you want to push the edit further, or just save the image.

Install

Install the pack once; that covers every node in it. ComfyUI Manager: search "FastTool". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Fengxiaoxiao-001/ComfyUI-FastTool.git

Then restart ComfyUI. No Python dependencies - pyproject.toml declares none; it builds on ComfyUI's own comfy.samplers, comfy.model_management, and SDXL VAE code.

Where people get burned

• It's VRAM-heavy by construction. Four UNet inputs per noise per step adds up fast. If you OOM, drop unet_batch_size to 1 first, then noise_samples, then n_steps. There's also a clear_cache_after_run toggle (default off) that frees the cache when the node finishes.
• It's SDXL-only. The conditioning uses SDXL's crop-coordinate encoding; don't feed it an SD1.5 or Flux setup and expect it to work.
• Empty or near-identical prompts log a warning (or raise, if identical). Give both prompts real content - the edit direction is the difference between them.
• The pack README describes more nodes than the current build registers (the VRAM CLIP Offloader it advertises is commented out in __init__.py). Trust what shows up in your node search, not the README.