ComfyUI LoRA Models: Verified Realistic, Speed & NSFW Downloads

Pick the right file first

A LoRA must match the model family it was trained for. Do not load an SD 1.5 LoRA into SDXL, or an SDXL LoRA into Flux. For resources with several architecture tabs, select the named version below instead of downloading whichever file CivitAI currently shows by default.

“Free to download” does not mean “unrestricted.” Read the displayed license before commercial use or redistribution. Flux.1 Dev derivatives on this page can carry the FLUX.1 [dev] Non-Commercial License.

Verified quick reference

The proposed F.2 Midjourney Flux.2 card is held and removed because CivitAI displays a FLUX.1 [dev] license notice for a Flux.2 file. Hyperrealistic is also held because version-pinned source metadata was unavailable. FLUX NSFW Unlock is removed because its account-gated page did not expose enough authenticated file, trigger, strength, and license data.

Detail controls

Detail Tweaker LoRA v1.0 — SD 1.x

The creator documents a weight range from -2 to 2: positive values increase detail and negative values reduce it. No trigger word is listed. Download the 36.11 MB SafeTensor and use it only with a compatible Stable Diffusion 1.x base.

Detail Tweaker XL v1.0 — SDXL

The SDXL version is a separate 217.87 MB file. The creator documents -3 to 3 and recommends starting at 1.5, then adjusting for the prompt and sampling steps. No trigger word is listed.

xl_more_art_full-v1 — SDXL 1.0

This 685.71 MB version is the exact file linked above. The creator recommends 0.8–1.0 and describes it as a general aesthetic, artistic, creative, and detail enhancer. Aerial is an optional keyword for aerial-photo training, not a required activation token.

FLUX Image Upgrader v0.3 — Flux.1 Dev

Choose FLUX v0.3, not adjacent SDXL, SD 1.5, or Illustrious versions. The official version lists strength 0.4, documents a broader 0.15–1.0 range, and lists aidmaimageupgrader. Its displayed license is non-commercial.

Add More Details v1.0 — SD 1.x

This separate 9.11 MB SD 1.x LoRA uses a creator-recommended 0.5–1.0 range. It overlaps with Detail Tweaker, so choose between them before stacking.

Speed LoRAs

LCM & TurboMix — Euler a.fix

The selected SDXL file is 12.36 MB. The creator documents weight around 1, CFG 2, and 6–8 steps. Version names correspond to samplers; this link opens Euler a.fix. Match the selected sampler to the downloaded version.

ByteDance SDXL-Lightning

Use the official ByteDance repository, not a CivitAI mirror. LoRA files are provided for 2, 4, and 8 steps; use the file matching the workflow's step count. Official ComfyUI instructions specify:

1. Put sdxl_lightning_Nstep_lora.safetensors in ComfyUI/models/loras/.
2. Use the official LoRA workflow.
3. Use Euler with sgm_uniform.
4. Match inference steps to the downloaded file and use CFG 0.

The project also publishes a 1-step model, but it is an experimental full/UNet checkpoint, not a 1-step LoRA.

Style and realism

Touch of Realism V2 — SDXL

The V2 file is 435.34 MB and uses touch-of-realismV2. CivitAI lists strength 0.6; the creator describes 0.5 as a subtle starting point. It is trained for SDXL.

Realistic Skin Texture — choose an architecture version

This CivitAI resource contains separate Z-Image, Flux.1 Dev, SDXL, SD 1.5, Pony, and Illustrious files. One file does not cover every architecture. Choose the base family you use, then verify that version's size, triggers, and license.

PhotorealTouch — choose an architecture version

PhotorealTouch publishes separate Z-Image, Pony, SDXL, SD 1.5, Illustrious, and Flux versions. The current default, Z - IMAGE V1.0, is not an SDXL file. Its 7–10 steps / CFG 2–3 guidance is scoped to Z-Image/Turbo; do not copy those settings to another architecture.

Why the held cards are absent

F.2 Midjourney V6.1 meets FLUX

The page artifact verified a pinned Flux.2 v0.5 file, but CivitAI displayed a FLUX.1 [dev] Non-Commercial notice on that Flux.2 version. Architecture and license labeling must agree before the card is safe to publish. The card is therefore removed rather than shipping the conflict as a footnote.

Hyperrealistic

The official CivitAI URL resolved, but the public source did not expose enough version-level metadata to verify the file, size, trigger, base model, compatibility, and license. The prior 435 MB and Hype4realistic claims are not retained.

FLUX NSFW Unlock

The official resource is account-gated. Its public response identified a Flux.1 LoRA, but did not expose enough authenticated metadata to verify the exact file size, trigger, license, strength, or the previous claim that it is required. Removing the card is safer than retaining a placeholder or unsupported technical fields.

How to evaluate a multi-version resource

CivitAI resources can place unrelated architecture files under one model page. Before downloading:

1. Open the version selector rather than accepting the default tab.
2. Match the displayed base model to the checkpoint you already use.
3. Record the exact version name and model-version URL.
4. Confirm the selected file's size, trigger words, creator-recommended strength, and displayed license.
5. Re-check whether settings in the description apply to every version or only the currently selected one.

A resource-level title or download count does not prove that one file works across every listed architecture. Realistic Skin Texture and PhotorealTouch remain in the catalog only as version selectors, not as universal files.

File size, runtime memory, and licenses

A LoRA's on-disk size is not its VRAM requirement. Runtime memory is dominated by the base checkpoint, text encoders, VAE, precision, resolution, batch size, attention backend, and offload strategy. That is why the refreshed model-card payload leaves vram blank rather than repeating unsupported per-card floors.

The same caution applies to availability and commercial use. A file can be downloadable without charge while still carrying non-commercial, attribution, redistribution, or derivative restrictions. Read the license shown for the exact selected version, not a neighboring architecture tab.

Install a LoRA in ComfyUI

Official ComfyUI documentation confirms this workflow:

1. Download the selected LoRA from its creator or official repository.
2. Put it in ComfyUI/models/loras/ (subfolders are supported).
3. Refresh or restart if it does not appear.
4. Add Load LoRA.
5. Connect the base model and clip, select lora_name, and set strength_model / strength_clip.
6. Multiple Load LoRA nodes can be chained, but there is no universal three-maximum or combined-strength rule. Follow each creator and change one adapter at a time.

For screenshots and a full workflow, see How to install and use LoRAs. For compatible base checkpoints use the Stable Diffusion model catalog.

Troubleshooting

The LoRA does not appear: confirm it is under ComfyUI/models/loras/, then refresh or restart. Subfolders are supported.

The workflow errors when loading: verify the LoRA's base architecture against the loaded checkpoint. An SDXL, Pony, Illustrious, Flux.1, Flux.2, Z-Image, or SD 1.5 label is not interchangeable.

The result is too strong or distorted: return to the creator's documented starting strength, remove other LoRAs, and change one adapter at a time. Do not rely on a universal combined-strength formula.

A speed LoRA produces poor output: match the selected file to its documented sampler, scheduler, CFG, and step count. SDXL-Lightning's 2-, 4-, and 8-step files must match the workflow.

No model was downloaded or run for this refresh. “Verified” refers to source-card metadata, not hands-on quality testing.