Live Nov 18 · Cohort 3Nov 18, 19:00 MSK — free hands-on AI agents live workshop · Live Cohort 3 starts Nov 23Join Nov 18 Live →Cohort 3 (Nov 23)
InfrastructureInsight · On RadarHabr AI→ original

sd.cpp instead of Forge Neo: nine models on a 16 GB GPU draw in 5–68 seconds

The author replaced the Python-based Forge Neo interface with sd.cpp, a single C++ executable for running models, and compared nine image generators on an RTX 4070 Ti SUPER with 16 GB of VRAM. Z-Image Turbo renders a 1024×1024 frame in 5 seconds and writes Russian text correctly. Moss, the best at people, takes 68 seconds. The "Fast" button speeds up the slow models by 1.5–2.4x, but complex scenes come out smeared. For speed, pick Z-Image Turbo; for people and text, pick Moss.

AI-processed from Habr AI; edited by Hamidun News
sd.cpp instead of Forge Neo: nine models on a 16 GB GPU draw in 5–68 seconds
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Fast modeMoss 68→28 s, Qwen-Image 30→16 s, Chroma 60→40 s
Key metric #2Flux.1 dev, Chroma1-HD and Anima cannot write Cyrillic
Key metric #3The 8 GB version usually takes as long as the 16 GB one

If you have ever generated images on your own GPU, you know the problem: Stable Diffusion Forge Neo and ComfyUI are huge Python applications with gigabytes of PyTorch and packages. After each update something breaks, and add-ons conflict with one another. The author of a Habr article walked away from all that: he moved his app Imagecaster to stable-diffusion.

cpp (sd.cpp for short), a single C++ executable with no Python. He then ran nine models on a home RTX 4070 Ti SUPER with 16 GB of VRAM.

The numbers in the headline are seconds per 1024×1024 image: the fastest model, Z-Image Turbo, finishes in 5 seconds, and the slowest, Moss, in 68. "16 GB" and "8 GB" are two variants of the same model with different levels of weight compression (quantization): the first is meant for a card with 16 GB of memory, the second for a card with 8 GB. The practical takeaway is simple: if you need speed, pick one model; if you need good people and accurate Russian text, pick another.

You can run both without wrestling with Python.

Why sd.cpp instead of Forge and ComfyUI

The app's server has gone from A1111 to Forge, then Forge Neo, and now runs on sd.cpp. Models come in GGUF format: each model has several compression levels, so you can pick a file that fits your VRAM. There are builds for AMD and Intel (via Vulkan) and for Macs with Apple Silicon (via Metal), and the server picks the right one itself. But the author tested only on NVIDIA.

He also dropped Ollama: the server now works directly with llama.cpp, and the author built model downloading and unloading into the server itself. The approach has downsides: new models show up in sd.cpp later than in ComfyUI, and it does not always understand fp8 files from ComfyUI. For that reason he had to use the full version of HiDream.

The test worked like this: five difficult prompts (a watchmaker with a note saying "Back at 13:60", musicians on a St. Petersburg rooftop, a city inside an oak, an anime scene titled "I'm late!", a 1998 dacha), with four seeds for each model version. Timing covers generation only, not model loading, averaged over 20 images.

Nine models: speed and Russian text

Z-Image Turbo — 5 seconds in 8 steps, photorealism, Russian text without errors. The 8 GB version (3.9 GB) runs as fast as the 16 GB version (6.6 GB).

Flux.3 Klein — 8 seconds, takes plain-text prompts, writes Russian correctly. The 8 GB version weighs 3.5 GB and is just as fast.

Krea-2 Turbo — aesthetic photos in 8 steps and 11 seconds, with correct Russian text too.

Anima — a model for anime and illustration that understands both danbooru tags and ordinary phrases. 19 seconds per image; a single 4.2 GB version suits both 8 GB and 16 GB cards. Photos and text are not its job.

Flux.1 dev — a classic that follows descriptions well, but draws letter-like scribbles instead of Cyrillic. 25 seconds on 16 GB.

Qwen-Image 2.1 — follows the prompt precisely and writes legible text, including in Russian. It is the only one of the nine that runs at CFG 6, so it takes the negative prompt (a list of what should not appear in the image) into account. 30 seconds, 16 in fast mode.

HiDream-O1 dev — everything comes out in an illustrated style, which is why the author shows anime with it. Large lettering works; small text turns into scribbles. The working bf16 version weighs 16.4 GB and does not fit entirely into 16 GB of VRAM: part of the weights sit in system RAM, and an image takes 30 seconds. The fp8 version misbehaved in sd.cpp: it ignored the prompt and added artifacts.

Chroma1-HD — an open model with 8.9 billion parameters based on FLUX.1 schnell. It is valued for its Apache 2.0 license (Flux.1 dev's license is non-commercial), the absence of a safety filter, and true CFG with a negative prompt. Its strength is a filmic, cinematic look. Downsides: 60 seconds per image and scribbles instead of Russian letters.

Moss (OpenMOSS, 5 billion parameters) — the author's favorite: it draws people better than any other, follows the description precisely, and writes Russian correctly. The price is time: 50 steps and 68 seconds per image.

Fast mode and what is wrong with it

The "Fast" button turns on step caching: the engine reuses computations from neighboring steps instead of calculating each one from scratch. The mode is available for four slow models: Moss and Chroma (dbcache), Qwen-Image and Flux.1 (easycache). The speedup ranges from 1.5x for Chroma to 2.4x for Moss: for example, Moss drops from 68 to 28 seconds, and Qwen-Image from 30 to 16. With Moss and Qwen-Image the scene barely changes at the same seed and only details shift, while Chroma noticeably rebuilds the frame.

There is no free speedup. On a simple image at low resolution, the difference is barely visible. On a complex scene at high resolution, the image looks like an over-compressed JPEG, and intricate areas turn to mush. So you should not enable the mode everywhere: whether it works for a given image is visible only in the result.

The second observation: quantization saves VRAM, not time. For most models the 8 GB version takes as long as the 16 GB version. The author measured both on a 16 GB card, but sees no reason why they would not work on 8 GB. If memory runs out, the server moves part of the model into system RAM, and generation gets slower.

What to choose

If you need speed and Russian text, take Z-Image Turbo or Flux.3 Klein. For beautiful photos, Krea-2 Turbo. For accuracy and text at a moderate time cost, Qwen-Image. For people and maximum quality, Moss. For anime, Anima. The ratings are subjective: the author says outright that he simply does not like Flux.1 and Chroma. But neither of them writes Cyrillic, which is worth considering if you are building a product for Russian-speaking users.

Story

Key Questions Answered

Why sd.cpp instead of Forge and ComfyUI?

The app's server has gone from A1111 to Forge, then Forge Neo, and now runs on sd.cpp. Models come in GGUF format: each model has several compression levels, so you can pick a file that fits your VRAM. There are builds for AMD and Intel (via Vulkan) and for Macs with Apple Silicon (via Metal), and the server picks the right one itself. But the author tested only on NVIDIA.

What are the key details on Nine models: speed and Russian text?

Z-Image Turbo — 5 seconds in 8 steps, photorealism, Russian text without errors. The 8 GB version (3.9 GB) runs as fast as the 16 GB version (6.6 GB). Flux.3 Klein — 8 seconds, takes plain-text prompts, writes Russian correctly. The 8 GB version weighs 3.5 GB and is just as fast. Krea-2 Turbo — aesthetic photos in 8 steps and 11 seconds, with correct Russian text too. Anima — a model for anime and illustration that understands both danbooru tags and ordinary phrases.

What are the key details on Fast mode and what is wrong with it?

The "Fast" button turns on step caching: the engine reuses computations from neighboring steps instead of calculating each one from scratch. The mode is available for four slow models: Moss and Chroma (dbcache), Qwen-Image and Flux.1 (easycache). The speedup ranges from 1.5x for Chroma to 2.4x for Moss: for example, Moss drops from 68 to 28 seconds, and Qwen-Image from 30 to 16. With Moss and Qwen-Image the scene barely changes at the same seed and only details shift, while Chroma noticeably rebuilds the frame.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…