How to use a LoRA in ComfyUI

1. 01 / processMatch

   Compatible checkpoint, LoRA and permissions

2. 02 / processConnect

   Loader between base MODEL/CLIP and generation

3. 03 / processCompare

   Fixed-seed baseline versus documented strengths

4. 04 / processKeep

   Outputs, workflow and dependency notes

Apply an existing LoRA to an image workflow

A LoRA changes parts of a compatible model's behaviour. It can adapt a style, subject or concept without replacing the whole checkpoint. This guide loads an existing LoRA, compares it against the same workflow without that adaptation, and saves both the image and recipe. Training your own LoRA is a separate task.

Our example prompt describes a cream robot wearing an orange jacket. It is an original practice prompt, not a demonstrated output or a trained character identity. Instructions and sources were checked on 11 October 2026; generation, visual quality, runtime and hosted billing were not tested.

Choose the route before downloading

Free local: ComfyUI and the built-in Load LoRA node can run on your own compatible hardware. You need a base checkpoint as well as a matching LoRA. An adaptation file on its own cannot generate an image. If you have not completed a basic image workflow, start with your first ComfyUI workflow.

Paid hosted: Comfy Cloud provides the editor and GPU. Its pricing page explicitly places custom model and LoRA imports on Creator, Pro and Team plans. Do not choose a cheaper plan on the assumption that every LoRA can be uploaded. Hosted importing is still subject to supported formats, sources and available models; confirm the files you need before paying.

Match the adaptation to the base

Check the publisher's model card for the LoRA's base family, recommended checkpoint, trigger words, suggested strength and usage licence. An SD1.5 LoRA belongs with an SD1.5-compatible checkpoint, not an SDXL, Flux or Wan base. A .safetensors extension describes the container, not compatibility. LoRAs within the same family can still depend on a particular checkpoint or training setup.

The official beginner tutorial uses dreamshaper_8.safetensors with blindbox_V1Mix.safetensors. Follow its model links if you want that exact documented example, and inspect each publisher's current permissions before use. We have not verified those files' redistribution rights, so our download includes neither. The tutorial is evidence for the loading arrangement, not proof that this pair will preserve our robot's identity.

Use one adaptation first. Adding several makes it harder to understand whether a failure comes from the base model, an incompatible file or competing adaptations. Our LoRA explanation covers what the method changes.

Install and connect it

1. Put the base checkpoint in ComfyUI/models/checkpoints and the LoRA in ComfyUI/models/loras. Windows Portable nests those folders under ComfyUI_windows_portable/ComfyUI. Desktop installations can use configured model directories; check the Storage panel instead of guessing the installation path.
2. Refresh the node/model definitions with R or restart. Confirm both files appear in their respective loader lists. Downloading an HTML webpage renamed to .safetensors is not a model installation.
3. Open a known working SD1.5 text-to-image graph. Add the built-in Load LoRA node, described in the official node reference.
4. Connect Load Checkpoint's MODEL and CLIP outputs into Load LoRA's matching inputs. Connect Load LoRA's MODEL output to KSampler's model input. Connect its CLIP output to both positive and negative CLIP Text Encode nodes.
5. Keep the original checkpoint's VAE connected to VAE Decode. Keep the encoded positive/negative conditioning, Empty Latent Image, KSampler, VAE Decode and Save Image chain. Do not leave KSampler connected directly to the old MODEL output: that bypasses the adaptation.

The upstream examples also show how adaptations sit between MODEL/CLIP loading and generation. No custom node pack is needed for this basic loader arrangement.

Make a useful comparison

1. Select the correct checkpoint and LoRA filenames. Copy any required trigger words from the adaptation's publisher; do not invent one from the filename.
2. Use this subject prompt, adding only the actual documented trigger words if needed:

small cream-coloured robot, orange jacket, standing beside a wooden desk, three-quarter view, plain slate background, soft studio lighting

3. Use text, watermark, duplicate subject as a simple negative prompt. Keep your working base workflow's sampler, steps, CFG and dimensions. For the SD1.5 starter, the inspected official JSON uses 512 × 512, batch one, 20 steps, CFG 8, euler, normal, denoise 1. These are a baseline, not a universal LoRA preset.
4. Choose seed 424242 and fixed seed control for the exercise. Set strength_model and strength_clip to 0 for the baseline run, then save its output. Keep the LoRA loaded so the graph and prompt remain otherwise unchanged.
5. Set both strengths to the publisher's recommended values and run again. If no strengths are documented, 0.6 for each is a proposed exploratory starting point, not a validated setting. Record it and compare before increasing further.
6. Check whether the intended style or subject actually changed, and whether important details worsened. A stronger style effect is not automatically a more useful image. If the result improves, save it; if it fails, preserve the failure and its settings too.

Save the image and the dependency record

Save the output through Save Image's image menu and confirm the file opens locally. The local node also saves under ComfyUI/output. Export the graph through Workflows → Export as robot-lora-comparison.json.

Keep a note of the exact base filename, LoRA filename, publisher URLs, file versions or hashes, trigger words, both strengths, seed and software versions. Those dependencies are not bundled inside the workflow JSON. Someone receiving only the JSON still needs compatible weights and permission to use them.

Use the hosted route

Open Comfy Cloud, check its available models and your plan's import eligibility, then import the matching checkpoint and LoRA from the supported publisher source if needed. Load your graph, reselect the cloud model entries and confirm the same MODEL/CLIP wiring. Run the baseline and adapted comparison separately, download each image and export the edited graph.

The subscription uses credits for active runtime; include both comparisons and unsuccessful attempts in any cost-per-usable-image calculation. No cost was observed for this guide. A pre-installed LoRA may avoid importing, but it still must match the base. Ordinary links here carry no active affiliate tracking.

Diagnose the result

• No visible effect: verify the loader's MODEL output reaches KSampler, its CLIP reaches the text encoders, strengths are non-zero and required trigger words are present.
• Shape mismatch or load error: check family and exact base compatibility before changing strengths. A lower weight cannot repair an incompatible architecture.
• Distorted result: return to the saved baseline; lower one strength or use the publisher's settings. Change one variable per comparison.
• Same character not retained: a style LoRA need not encode character identity. Use the character-consistency guide for that separate task.

The download supplies prompts, a comparison checklist and reproduction notes. It contains no weights or fabricated workflow JSON. Our intended comparison remains documentation-based until real baseline and adapted exports are retained.

small cream-coloured robot, orange jacket, standing beside a wooden desk, three-quarter view, plain slate background, soft studio lighting