Doom Translate

This is the node that quietly breaks the pack's "no extra dependencies" promise, and it's worth it: Doom Translate runs a real machine-translation model locally - Tencent's HY-MT 1.5 family - so you can write prompts in your own language and translate them to English (or any of ~36 target languages) without sending your text to a cloud API. No key, no account, no rate limits. The name is honest: it translates, on your hardware, and the model downloads from HuggingFace on first use.

It's aimed squarely at the pack's Russian-speaking audience, but the use case is universal: if English isn't your first language, or you keep a prompt notebook in your native tongue, this is the node that converts your ideas into text the model was trained on. Most modern encoders - Flux, Krea 2, H3 - were trained with English-dominant captions, so writing in your language then translating is genuinely better than feeding the model your native prose directly.

How it works

The node loads a transformers model from the Tencent HuggingFace repos (tencent/HY-MT1.5-1.8B-FP8 is the default), tokenizes your text, generates up to max_new_tokens, and returns the translation. There are four model choices - the 1.8B and 7B sizes, each in plain and FP8 versions. FP8 is the sensible default: a fraction of the VRAM for essentially the same output on translation.

The inputs that matter

• text - what you want translated. Multiline.
• target_language - from English to Vietnamese, ~36 languages in the list. The model picks the output language from this.
• model_name - HY-MT1.5-1.8B-FP8 (default), 1.8B, 7B-FP8, or 7B. Bigger is more fluent but eats VRAM.
• dtype_mode - auto / bfloat16 / float16 / float32. bfloat16 default.
• device_map_mode - auto / cpu / cuda:0. auto spreads across what you have.
• keep_model_loaded - keep the model cached in VRAM between runs. Off by default, and off is usually right: it's a translation model, not your sampler, and a few seconds of reload beats a permanently pinned chunk of VRAM.
• trust_remote_code - the tooltip says it plainly: "Allow execution of custom code from the model repository." Leave it off unless you have a specific reason.
• max_new_tokens - cap on generated tokens. 2048 default.
• seed - for reproducible sampling.

Output is a single translated_text string, ready to feed a sampler or a prompt builder.

Setup - read this, it's the one that differs

The pack install is the usual:

cd ComfyUI/custom_nodes
git clone https://github.com/PeterMikhai/Doom_Flux_NodePack

But this node needs transformers and huggingface_hub, which ComfyUI does not ship by default. The imports are lazy - the pack loads fine without them - but the first time you run Doom Translate it will fail with an ImportError unless you install them:

pip install transformers accelerate huggingface_hub

Then restart ComfyUI. This is the one place the README's "everything runs on ComfyUI's own stack" claim falls over.

Where people get burned

• The first run downloads a model. The 1.8B FP8 is a ~1–2 GB download from HuggingFace; the 7B is several GB. First run will look frozen while it fetches. It's downloading, not hung.
• VRAM. Even FP8, a 7B model is a real chunk of memory sitting next to a 12B diffusion model. Prefer the 1.8B unless you need the fluency, and keep keep_model_loaded off.
• trust_remote_code should stay false. It exists for repos that need custom code to run. A translation model from a major vendor doesn't. Enabling it on an unknown repo is how you get malware - this node's own tooltip is the warning.
• Translation quality follows model size. 1.8B is fine for prompt-length text; long prose and tricky idioms favor the 7B. For prompt work, the small model is genuinely enough.

It's the odd node in a pack that's otherwise all sampling and saving - but if you're building prompts in a non-English language, it's the one that makes the rest of the pack actually work for you.