Models, LoRA & Extensions

Stable Diffusion LoRA Weights Explained

PhoneDiffusion
Stable Diffusion LoRA Weights Explained

Models & LoRA

A LoRA weight (also called strength or multiplier) sets how much of the LoRA's learned change is added to the base Stable Diffusion model: 0 means the base model alone and 1 means the LoRA at full trained strength. Start with the creator's recommended weight, usually between 0 and 1, and adjust it with a fixed prompt and seed.

What do LoRA weights mean? The short answer

A LoRA weight is a multiplier on the LoRA's effect. A LoRA stores a small learned change to the base model, and the weight decides how much of that change is applied when you generate. Hugging Face Diffusers puts it plainly: a weight of 0 is the same as using only the base model, and a weight of 1 fully uses the LoRA.

Tools give the same number different names: multiplier in the AUTOMATIC1111 <lora:filename:multiplier> tag, strength_model and strength_clip on ComfyUI's LoraLoader node, and adapter_weights or scale in Diffusers. The weight is not a quality score. Too low and the style or subject barely shows; too high and the LoRA can take over composition, color, and anatomy.

What LoRA weight should you use?

Start with the weight the LoRA's creator recommends, if the model card gives one. The AUTOMATIC1111 wiki describes the multiplier as a number generally from 0 to 1. Its default when you click a LoRA card is 1.0 (the ‘Default multiplier for extra networks’ setting, a 0 to 2 slider), unless you saved a Preferred weight in that card's metadata editor.

Then test instead of guessing. Keep the prompt, seed, sampler, steps, CFG, and size fixed and generate the same image at a few weights, for example 0, 0.5, 0.75, and 1. The image at 0 shows the base model; the steps upward show what the LoRA adds. Pick the lowest weight where the concept is clear, because lower weights leave more room for the prompt and for other LoRAs.

If the concept only appears near full strength, check the basics before raising the weight further: the LoRA must match the checkpoint's model family (SD 1.5, SDXL, and so on), and many LoRAs need a trigger word in the prompt. A stronger weight does not repair a mismatched LoRA.

How LoRA weight works in AUTOMATIC1111

In AUTOMATIC1111 you set the weight in the prompt tag: <lora:filename:0.7> applies the LoRA at 0.7. The tag only switches the LoRA on and is removed from the prompt text before generation. The wiki notes that a LoRA cannot go in the negative prompt, and that a batch with different prompts uses only the first prompt's LoRA.

The tag can also split the weight between the two parts a LoRA usually patches. AUTOMATIC1111's source reads the second value as the text-encoder multiplier and an optional third value, or unet=, as the UNet multiplier, as in <lora:filename:te=0.5:unet=1>. If you give only one number, both parts use it.

Model strength vs CLIP strength in ComfyUI

ComfyUI's LoraLoader node has two weights. strength_model changes how strongly the LoRA modifies the diffusion model, the part that denoises the image. strength_clip changes how strongly it modifies the CLIP text encoder, the part that turns the prompt, including trigger words, into guidance. Both default to 1.0.

Lowering strength_model weakens the LoRA's look in the image. Lowering strength_clip changes how the LoRA affects the way the prompt is read. If you only want the image part, the LoraLoaderModelOnly node applies the LoRA to the diffusion model alone. Setting both strengths to 0 skips the LoRA entirely, which is a quick way to make a base-model comparison in the same graph.

To stack several LoRAs, chain LoraLoader nodes; each keeps its own pair of strengths. Combining multiple LoRAs covers that workflow.

PhoneDiffusion generates on-device on iPhone or iPad, with optional cloud models. Download on the App Store →

LoRA weights in Diffusers, including per-block scales

In Diffusers, load a LoRA with load_lora_weights and an adapter_name, then set its weight with set_adapters, for example adapter_weights=0.7. The documentation warns that set_adapters only scales attention weights: other LoRA layers, such as ResNets or down and upsamplers, stay at a scale of 1.0.

Instead of one number you can pass a dictionary that scales each part separately. The Diffusers SDXL example sets both text encoders to 0.5, the UNet's down blocks to 0.9, and individual up blocks to different values; any block you leave out stays at 1.0. When you fuse a LoRA into the model with fuse_lora, set the strength with its lora_scale argument, because a scale passed at generation time no longer applies.

Negative weights and dynamic LoRA weights

A weight can be negative in some tools. ComfyUI's LoraLoader accepts strengths from -100 to 100 and its tooltip says the value can be negative. A negative weight applies the LoRA's learned change in the opposite direction instead of adding it, which can push the image away from what the LoRA learned; results vary by LoRA, so test it like any other weight.

Dynamic LoRA weights change during sampling instead of staying fixed. Diffusers documents scale scheduling with a step callback: its example starts a character LoRA at 1.5, lowers it to 0.7 over the first 20 of 30 steps to set up the character, then holds 0.2 so the LoRA does not spread into parts of the image it was not trained on. Diffusers notes that changing the scale during sampling gives better control over composition and layout, because some steps benefit from a higher or lower scale.

AUTOMATIC1111 has no built-in schedule, but the LoRA Control extension (sd-webui-loractl) adds one with weight@step lists. Its README example <lora:name:0@0,1@0.5,0@1> starts at 0, reaches 1.0 halfway, and returns to 0 at the end; steps from 0 to 1 are read as a share of the pass, te= and unet= schedule the text encoder and UNet separately, and hr= sets the weight for the high-res pass.

Why 1.0 is not the same strength for every LoRA

A LoRA file also carries its own scale from training. The LoRA method keeps the base weights frozen and learns small low-rank matrices; their rank (dimension) and an alpha value set how large the stored change is. Hugging Face PEFT scales a LoRA by lora_alpha divided by rank by default, and AUTOMATIC1111's loader multiplies the update by alpha / dim before applying your multiplier.

So 1.0 means ‘as trained’, not a standard amount of effect. Two LoRAs at the same weight can differ in strength, which is why a creator's recommended weight is worth reading and why each LoRA needs its own test rather than one weight for everything.

PhoneDiffusion availability and article scope

PhoneDiffusion runs curated Core ML model packs (SD 1.5, SD 2.1, and SDXL) on supported iPhone and iPad devices with iOS 17 or later; local generation is private and works offline once a pack is downloaded. It also offers optional cloud models, such as FLUX and Seedream, on paid monthly or yearly plans; with those, the text prompt is sent online to generate the image.

The app does not provide a general production importer for arbitrary LoRA files, checkpoints, ControlNet files, or models from Civitai or Hugging Face, and it has no LoRA weight setting. Its Transformation slider (image-to-image) and Replacement strength (inpainting) control how much of a source image changes, not a LoRA weight.

The multipliers, per-block scales, negative weights, and schedules above describe AUTOMATIC1111, the LoRA Control extension, ComfyUI, and Diffusers. Use one of those desktop or server tools when a specific LoRA and weight matter to your result.

Field notes

  • A LoRA weight multiplies the LoRA's learned change: 0 is the base model, 1 is the LoRA at full trained strength.
  • Start with the creator's recommendation, usually between 0 and 1, and keep the lowest weight where the concept shows.
  • AUTOMATIC1111 and ComfyUI can weight the text encoder and the image model separately; Diffusers can also scale individual blocks.
  • Weights can be negative or scheduled over steps in some tools, and 1.0 means ‘as trained’, so equal weights are not equal strength across LoRAs.

Try it on your phone

  1. Confirm the LoRA matches the checkpoint's model family and add its trigger words.
  2. Generate a base image with a fixed prompt, seed, sampler, steps, CFG, and size.
  3. Repeat it with the LoRA at a few weights, such as 0.5, 0.75, and 1.
  4. Keep the lowest weight where the concept is clear; adjust text-encoder and model strength separately if only one part is too strong.
  5. Save the LoRA file version and its weights with the image.

Sources and verification

Primary and maintainer documentation checked for this guide. Product-specific details were reviewed against the current PhoneDiffusion scope.