Improved reForge is aimed at enhancing performance and optimization. Feel free to open an issue if you have any optimization ideas.

There are also some minor config adjustments to improve generation speed, while still staying within a safe range without reducing image quality.

---

• Removed some unecessary features.
• Improved lora-block-weight model accuracy.
• Seed when generating multiple images with batch count now randomized and no longer incremental.
• Added Sage Attention and Flash Attention. It's not really a big performance gain, but there is a noticeable speed improvement, around 0.7-0.8 it/s.
  SageAttention and FlashAttention setup
  Refer: lllyasviel/stable-diffusion-webui-forge#2866
• Now you can drop an image and load the metadata directly in the txt2img tab, without needing to go through PNG Info → Send to txt2img.
• Added Batch Diversity extension.
  • Purpose: helps batches generated from the same prompt produce more varied angles, framing, and composition.
• Added Inline Negative Prompt extension.
  • Purpose: allows per-line negative tags in batch prompt-list workflows by moving (n:...) markers from the positive prompt to the matching negative prompt.
  • Example: from below, low angle, (n:from below) becomes positive low angle and negative <static negative>, from below.
• Added Inline Resolution extension.
  • Purpose: allows per-line width/height in batch prompt-list workflows via (r:width, height) markers; lines without a marker fall back to the WebUI base resolution.
  • Example: wide shot, city, (r:1024, 1380) renders that line at 1024x1380; the marker is removed before generation.

Using Character and Style LoRAs

impForge provides native SDXL LoRA role splitting to reduce interference between character and style LoRAs. It is intended for SDXL models, especially IllustriousXL.

Use role=style for a style LoRA and role=char for a character LoRA:

<lora:style:0.9:0.9:role=style>
<lora:character:0.95:0.95:role=char>

A practical starting point when combining both LoRAs is:

<lora:style:0.8:0.8:role=style>
<lora:character:1.0:1.0:role=char_strong>

Available roles:

Role	Purpose
style	Preserves the style while reducing character and composition leakage.
style_pure	Applies stricter style isolation. Use it when the style LoRA strongly changes the character.
char	Default character profile for identity, hair, eyes, clothing, and accessories.
char_strong	Strengthens character identity when details disappear after adding a style LoRA.
char_max	Maximum identity profile. Use cautiously because it may introduce artifacts or overpower the style.

Start with style and char. Move to char_strong only when character details remain too weak. Lower the style strength before trying char_max.

Custom and Legacy Block Weights

Existing 12-value lbw= prompts remain supported:

<lora:style:0.9:0.9:lbw=0,0,0,0,0,0,0,1,1,1,1,1>
<lora:character:0.95:0.95:lbw=0,0,0,0,0,0,0.9,0.9,0.9,0,0,0>

Both 12-value legacy lists and 13-value extended lists are accepted. Use either role= or lbw= for one LoRA. If both are provided, lbw= takes precedence.

Malformed lists, non-finite values, and unknown roles produce an error instead of silently falling back to another profile.

Effective Weight

A role= preset multiplies the strength written in the prompt, so the two are not the same number. role=char at 0.9 used to apply as 1.03 on some blocks, past anything the LoRA was trained at, which is what let a character LoRA impose its dataset's habitual posing over the prompt.

Effective weight is now capped at 1.0, so the number in the prompt is the number applied. Turn off Cap effective LoRA block weight at 1.0 in Settings → Extra Networks to reproduce images made before this change.

The generation metadata records what actually ran, for example LoRA effective peak: char:1.00@IN04.

Step Windows for Pose Control

If poses always come out generic and complex poses never work, the character LoRA is usually competing with the prompt over the composition. Composition is decided in the first few steps; identity and detail are built afterwards. Delay the LoRA instead of weakening it:

<lora:character:0.9:0.9:role=char:start=0.15>

The first 15% of the run forms the composition from the prompt alone, then the LoRA applies at full strength for the rest, so identity is not sacrificed. Start at 0.15 and adjust from there: lower it if identity weakens, raise it if the pose still does not respond.

Bounds are written either as a fraction of the run (start=0.15) or as whole steps (start=4). Fractions are recommended, because hires fix resamples at a different step count and a fraction scales to it while a fixed step index does not. stop= and step=start-stop are also accepted; both bounds of one tag must use the same form.

Behaviour across passes:

• Hires fix — fractional windows scale to the pass. Absolute windows are skipped there and the LoRA runs at full strength, since a step index calibrated for the base pass would delay it again during the pass that builds final detail.
• ADetailer — windows are skipped entirely. Face inpainting does not decide composition, so muting the character LoRA there would only remove identity from the fix.
• Lowvram — windows are declined with a warning, because weights are patched per forward pass instead of baked once.

Step windows apply to UNet patches only; text-encoder conditioning is computed before sampling starts and keeps its requested strength. Crossing a window boundary re-bakes the affected weights, which shows up as a brief pause on that step.

Batch Restriction

Every prompt in one processing batch must use the same LoRA names, strengths, and roles. Run prompts with different LoRA stacks as separate jobs. This prevents later prompts from silently using the first prompt's LoRA configuration.

Role splitting improves isolation but cannot recover details that the LoRA did not learn reliably. Exact tattoos, small symbols, and asymmetric accessories may still require higher resolution, regional conditioning, inpainting, or LoRA retraining. More detailed documentation is available in SDXL LoRA Role Split.