Imagecaster: local models running on 16 GB learn to write in Russian and try on new looks

Six fingers and mangled Russian text are the defining stereotypes of local image generation. The creator of Imagecaster, formerly SD Studio, checked how well those stereotypes held up six months after the first article. Testing ran on an RTX 4070 Ti Super with 16 GB of VRAM.

Which models made the cut

In May, the app supported only SD 1.5, SDXL and Pony. They require tag-based prompts and struggle with complex scenes such as “right hand on a sign, left hand adjusting hair.” The core lineup now consists of Flux 1, Z-Image Turbo, Moss-Image and Flux 2 Klein. They understand ordinary sentences but require 12–16 GB of VRAM.

Each model has its niche. Flux 1 follows long descriptions closely. Z-Image Turbo delivers photorealism in 8 steps instead of 20–50, making it useful for exploring variations. Moss-Image renders a described scene as well as Flux 1 and supports editing with Russian instructions: “replace the trousers with a skirt” changes the clothing while preserving the person, pose and background. Klein handles post-processing and remixes that preserve the face.

The author tried Qwen-Image 2.1 and Krea 2, then removed them. The combination of Z-Image, Moss and Klein covers the same tasks with less overhead and greater speed.

The Russian-text test

The prompt: a young woman at a market holding a sign reading “Kartoshka, nedorogo” (“Potatoes, cheap”). The results:

• Moss: 50 steps, around 2 minutes per image, 5 out of 10 without errors. Mistakes were minor, such as “nedoroggo.”
• Z-Image Turbo: around 15 seconds per image, 6.5 out of 10. Errors included “Kартosha” rendered as “Kартosha” in transliteration: “Kартosha”.
• Flux 1: around a minute per image, 0 out of 5.

According to Claude, the explanation lies in the text encoder rather than quantization. Flux 1 reads prompts through T5, trained primarily on English. Z-Image and Moss use multilingual encoders from the Qwen family. The samples are small, at 5–10 images, so these results are a guide rather than a benchmark.

An AI stylist that keeps images out of the cloud

A local LLM that fits alongside image generation within 16 GB is underpowered for the stylist role. The workflow therefore works like this: a local vision model describes the photo, a capable agent (Claude Code or Codex) receives the text via MCP, and the app generates the image. The image never leaves the computer.

The trade-off is that the agent cannot see the photo itself. Conversation can help compensate: users can explain their tastes and anything the vision model missed. In one instance, the stylist assumed the picture showed the author's wife, even though the woman was AI-generated.

Faces pose a separate problem: Klein redraws the entire image. The eyes, nose and lips are therefore transferred from the original using landmarks, and the likeness is checked. If the face does not match, generation is retried, with up to three attempts. Across a set of seven variations with different hairstyles and hair colors, the facial-difference metric stayed within 0.013–0.036, in the green zone.

What this means in practice

If you are building a product with local image generation and a Russian interface, look at the model's encoder as well as its size. Only models with Qwen encoders are really suitable for Cyrillic lettering, and even then you will need to generate a batch of variations. A “cloud brain + local eyes and hands via MCP” setup is a workable way to preserve privacy when a small local LLM cannot handle the task.

The limitations are clear: heavy models take up to a couple of minutes, conversational editing requires an AI agent, and the graphics card needs at least 16 GB of VRAM. The imagecaster-server installer simplifies setup: it installs Forge Neo, Ollama if needed, and the required versions of Python and git in its own folder.