Red Moon Miranda Gray

Getting Red Moon Miranda Gray Running in ComfyUI

Red Moon Miranda Gray is a character-focused LoRA and checkpoint combo that works decently for consistent portraiture, but it needs some specific setup to actually produce usable results. Most people just drop it in and wonder why the outputs look washed out or the face doesn't match the reference. Here is what actually works.

Red Moon Miranda Gray Setup

You need to place the base checkpoint in your models/checkpoints folder and the LoRA in your models/loras folder. The model was trained primarily on SDXL architecture, so putting it in an SD 1.5 workflow will just give you noise and bad proportions. I know because I tried that first. The checkpoint alone gives you a grayish skin tone baseline and muted color palette. The LoRA adds facial structure consistency and the specific expression quality that the name implies. Use them together at these weights:

• Checkpoint: Red Moon Miranda Gray v1.4 (or later if available)
• LoRA weight: 0.8 to 1.0
• Clip skip: 2

That clip skip value matters more than most guides mention. The model's training data used CLIP token weighting, and leaving skip at default (1) makes the faces look over-smoothed and plastic. Setting it to 2 restores some of the texture detail that gets lost otherwise. For prompts, keep the descriptor simple. The model responds well to basic appearance tags like woman, gray eyes, pale skin, red moon background but any more specific than that and the sampler starts fighting the LoRA's internal conditioning. I learned this after spending three hours trying to get a specific hand pose and ending up with fused fingers and a completely wrong face every time. Sampling settings that actually work: DPM++ 2M Karras, 30 to 40 steps, resolution 1024x1024 or 896x1152 for portrait orientation. Going below 28 steps produces visible artifacting in the face region. Going above 50 steps doesn't improve quality, it just introduces over-rendered skin texture that looks like digital clay. The sweet spot is narrow but it's there.

One edge case I ran into repeatedly: when using img2img with this model, the input image needs to be close to the target composition or the face reconstruction breaks down. I spent a full week debugging why my portrait generations would occasionally produce a perfectly good face and then suddenly degrade into a smeared mess. The issue was that my input images had different lighting angles than what the model expected. The fix was using a controlnet with openpose only, no depth or normal maps, and feeding it a reference image with frontal or three-quarter lighting. That eliminated the random degradation entirely. Another thing nobody warns you about: batch size greater than 1 causes memory issues on cards under 16GB VRAM. The checkpoint alone is about 6.5GB. Add the UNet and text encoders and you are pushing close to 14GB on a single 12GB card. I had to downgrade to batch size 1 and accept the slower throughput rather than deal with OOM errors mid-generation. It is a bottleneck if you need volume output, but the quality tradeoff is worth it for single-image work. If you are trying to use this model for full-body shots, honestly just switch to a different checkpoint. Red Moon Miranda Gray is optimized for head and shoulders framing. Full body attempts result in distorted legs and inconsistent clothing rendering. I tested this across forty-five generations before accepting it was a limitation of the training data, not my prompt wording. The model simply was not trained on full-body compositions and no amount of prompting will fix that.

Download the files from Civitai under the Red Moon Miranda Gray listing. Make sure you verify the model hash if you care about integrity. There are a few repackaged versions floating around that strip the metadata and sometimes include unwanted additional LoRAs in the download bundle. The official release is clean, the reuploads are not.