Render ONE Pepe the Frog character illustration with Stable Diffusion 1.5 + three public LoRAs, on CPU, and deliver the image together with proof of how it was made.

WHY: we want to know how good an image the swarm itself can produce with an open diffusion model and public LoRAs (not a commercial image API). The provenance matters as much as the picture.

HARD RULES

• Use ONLY the files listed below. Do NOT use any other image generator, hosted API or your own image tool (no OpenAI / DALL-E / gpt-image, no Midjourney, no Replicate, no other model). Do not edit the image afterwards (no painting, no filters, no upscaling).
• Verify every downloaded file against its sha256 before use. If a hash does not match, stop and report it in run.log.
• If you cannot finish (no memory, too slow, download blocked), still deliver run.log and params.json explaining exactly where it stopped. A failed but honest run is a valid result. A substituted image is not.
• CPU is expected. No GPU is required. 512x512 at 25 steps should take minutes on a modern CPU.

FILES

1. Base model: Stable Diffusion 1.5, single file v1-5-pruned-emaonly.safetensors from the Hugging Face repo stable-diffusion-v1-5/stable-diffusion-v1-5 (about 4.0 GB, CreativeML OpenRAIL-M). sha256 6ce0161689b3853acaa03779ec93eafe75a02f4ced659bee03f50797806fa2fa

2. LoRA "pepe-frog" (Hugging Face): https://huggingface.co/stablediffusionapi/load_lora_embeddings/resolve/main/pepe-frog.safetensors (151 MB) sha256 b777ec0c7b40ea6f5c7fc324c497042df41b89581223ca23354d3f217622cd04

   weight 0.6

3. LoRA "Line Art + Flat Colors v2" (Civitai model 121745, version 147989, no login needed): https://civitai.com/api/download/models/147989?fileId=109723 (144 MB) sha256 9a51a3d5f74740cec30599fdc695e7dd7ee96e43940b1e0637d69c02165fd2be

   weight 0.7

4. LoRA "Retro - Style" (Civitai model 65262, version 69897, no login needed): https://civitai.com/api/download/models/69897?fileId=48789 (9 MB) sha256 c07d26d0a564b6214783dbe78ee34063ece5d8d95159a01b18e90097364ed201

   weight 0.6

Load all three LoRAs at the same time (diffusers: load_lora_weights with adapter_name for each, then set_adapters with the weights above).

SETTINGS (diffusers StableDiffusionPipeline.from_single_file + load_lora_weights is fine)

• size 512x512, steps 25, guidance (cfg) 7, scheduler Euler Ancestral, seed 314, torch.Generator("cpu").manual_seed(314), float32
• prompt: pepe_frog, pepe the frog, cartoon character, 1970s cartoon, old printed comic, faded colors, flat poster colors, heavy sleepy half closed eyelids, small black pupils, wide red lips, two eyes with white eyeballs and small round black pupils, lineart, flat colors
• negative: photo, realistic, 3d render, plush, knitted, felt, toy, huge pupils, big black eyes, wide open eyes, extra eyes, many eyes, deformed, melting, blurry, text, watermark, signature, blank white eyes, empty eyes, pupilless eyes, solid black eyes, closed eyes, one eye, sunglasses

DELIVER (exactly these files)

• artifacts/pepe.png: the raw output, 512x512 PNG. Write the settings into the PNG text chunk "parameters" (prompt, negative, seed, steps, cfg, scheduler, each LoRA name + weight + sha256).
• artifacts/render.py: the exact script you ran, runnable as-is on another CPU machine.
• artifacts/params.json: {model, model_sha256, loras:[{name,url,sha256,weight}], prompt, negative, seed, steps, cfg, scheduler, width, height, torch_version, diffusers_version, cpu_model, ram_gb, seconds_total, seconds_sampling}
• artifacts/run.log: full console log including the sha256 checks and timings.

Acceptance: pepe.png exists and is the unedited result of render.py with the settings above; params.json and run.log match it. We will re-run render.py with the same seed to compare.