Ee General José Artigas

Getting Historical Figure Imagery Right: The José Artigas Challenge

Generating convincing portraits of 19th-century South American military leaders through diffusion models is genuinely harder than most people assume. You might think you just prompt "José Artigas general" and get a usable result, but the training data for this specific figure is patchy at best. The model has seen thousands of generic cavalries and colonial soldiers, maybe a handful of Uruguayan historical references, and your output will reflect that uneven coverage unless you know exactly where it breaks down. The "ee" tag in model names on CivitAI typically refers to models built on early diffusion architectures or specific fine-tune lineages. A model labeled with "general josé artigas" is usually a checkpoint or LoRA trained to produce period-accurate depictions of the Oriental leader. The quality varies wildly between versions. Some are decent at rendering the poncho, the feathered headband, and the gauntleted hand on a sword. Others just slap a mustache on a generic dark-skinned warrior and call it a day. I've personally dealt with a checkpoint that claimed high Artigas fidelity but consistently rendered him with anachronistic 20th-century uniform cuts, including epaulettes that didn't exist in the 1815-1820 period. The fix wasn't a better prompt — it was switching to a different LoRA and combining it with a VAE that preserved fabric texture detail instead of smoothing the poncho into featureless orange blob.

Setting Up the Right Model

Start by checking which base the model uses. SD 1.5-based checkpoints still dominate historical figure generation because they have the most supporting LoRAs and embedding packs. SDXL versions exist but tend to over-smooth period clothing details and make faces look too modern. If you're working in WebUI or ComfyUI, download the checkpoint from a reputable source and verify the file hash. For best results, pair the main model with these supplementary tools:

• A face refinement LoRA trained specifically on historical Latin American portraits — generic face LoRAs will give you an anachronistic modern aesthetic
• A clothing/textile embedding focused on 19th-century Gaucho and Oriental regional wear
• A VAE like freshvae or megavae — Artigas imagery fails hard without proper color reproduction in the reds and earth tones of the traditional poncho

Building Prompts That Actually Work

The most common mistake is assuming "José Artigas" alone is sufficient. In my testing, the model treats that as a weak positive and defaults to generic gaucho imagery. You need to reinforce specific visual anchors: Effective prompt structure: "General José Artigas, Uruguayan national hero, 1815, wearing traditional oriental poncho, feathered headband with crimson and white plumes, long dark hair, weathered face, standing beside a criollo horse, sword at hip, rural Pampas landscape, oil painting style, detailed fabric texture, historical accuracy, dramatic side lighting, 19th century military portrait"

Negative prompt essentials: "modern clothing, anachronistic, western uniform, epaulettes, steel helmet, modern boots, clean skin, youthful face, anime, cartoon, watermark, blurry, low quality, bad anatomy, extra fingers, missing limbs"

The Artifact Problem Nobody Talks About

Models hallucinate heavily on hands and weapons when generating Artigas imagery. This isn't a model defect — it's a data problem. There simply aren't enough high-quality reference images of Artigas holding a sword or resting a hand on a horse in the training set. I encountered this when a client needed a poster-quality image for a museum exhibition. Every output had either fused fingers, a sword melting into the horse's flank, or the poncho seamlessly blending into the saddle. The workaround I ended up using was a three-step process:

1. Generate the base image at a lower resolution (512x768) with a focus on correct pose and costume
2. Use inpainting to rebuild the hands and sword separately with tight masks
3. Upscale to final resolution using a dedicated details enhancer LoRA rather than a generic upscaler

This cut my failure rate from about 60 percent to roughly 15 percent across a batch of 20 generations. Don't generate at 1024x1024 expecting period accuracy. The model was primarily trained on smaller datasets for historical figures, and pushing resolution without additional refinement just amplifies artifacts. Start at 512x768 or 640x800, get the composition and costume right, then upscale in stages. A single aggressive upscale pass creates more problems than it solves. Use aCFG guidance between 5 and 7 for historical figures. Higher CFG values look sharp initially but produce oversaturated colors and that characteristic diffusion model wax-figure sheen on faces. Lower values (around 4) sometimes yield better skin texture and more natural fabric rendering, though you may need a second pass for definition.

When the Model Simply Can't Help

There are cases where no amount of prompting or post-processing will give you a reliable Artigas portrait. If the model version you're using has minimal Uruguayan historical training data, it will always default to a mestizo or generic South American warrior archetype. The poncho color might be wrong, the facial features too homogeneous, the setting anachronistic. In these situations, the honest recommendation is to either find a community-trimmed variant with better regional data or use a reference image pipeline with IP-Adapter to inject an actual historical painting as a structural guide rather than relying purely on text prompts.