Fat Mex Content Sampler

Type a prompt, pick a resolution, hit run, get an image. That's the whole deal with Fat Mex Content Sampler - it collapses EmptyLatentImage + CLIPTextEncode (pos) + CLIPTextEncode (neg) + KSampler + VAEDecode into one node. If you've ever dragged out that chain for the hundredth time, you know exactly why this exists. It's built for crank-out-content workflows where the graph shouldn't get in the way.

Wire it after a Fat Mex Model Loader and it's the simplest text-to-image you'll use all week.

How it works

The node does what its five raw equivalents do, but it borrows its sampling settings from the preset system. Every model family in this pack has known-good defaults baked in - steps, cfg, sampler, scheduler - so you don't have to remember that Qwen 2512 wants res_multistep or that Chroma HD runs at cfg 0.15. Set preset_hint to match the model you loaded and it picks those defaults automatically; auto falls back to Klein 9B-style settings (euler / beta / 6 steps / cfg 1.0).

A couple of mechanical details that matter:

• Resolution presets cover the usual 1024-based portrait/landscape ratios, plus Custom (which activates custom_width/custom_height).
• The latent is auto-shaped for the model. It starts as a standard 4-channel latent and fix_empty_latent_channels adjusts it for the newer 16-channel models (Flux 2, Qwen Image, Z-Image). You don't pick a latent format - that's handled.
• steps, cfg, sampler_name, scheduler are overrides, not defaults. 0 / auto means "use the preset," which is how you leave them for most people.

The inputs that matter

• positive_prompt and negative_prompt - the obvious ones. Keep in mind these are modern distilled models at low cfg: negative prompts are nearly inert at cfg 1.0, so write constraints as presence in the positive ("clean studio background") rather than praying the negative saves you.
• resolution - the 8 presets cover the standard ratios.
• seed - 0 gives you a fresh roll each time; set a value to reproduce.
• batch_size - how many images in one pass, and denoise - leave at 1.0 for full generation.

Outputs are images (the decoded IMAGE you preview and save) and latent (the raw latent, useful if you want to chain a second pass or an upscale without re-encoding from pixels).

Install

cd ComfyUI/custom_nodes
git clone https://github.com/FatMex/ComfyUI-FatMex-Nodes.git FatMex-Nodes
pip install -r FatMex-Nodes/requirements.txt

Or search "Fat Mex Nodes" in ComfyUI Manager and restart. The node itself needs nothing extra - the work is all in the model files.

Where people get burned

• The sampler doesn't know what you loaded. preset_hint defaults to auto, which assumes Klein 9B. If you loaded Qwen Image 2512 in the Model Loader and leave it on auto, you'll sample with Klein's defaults. Usually fine, occasionally weird. Get in the habit of matching preset_hint to the preset you picked in the loader - they're separate dropdowns and nothing links them.
• "It can't find my model." That error comes from the Model Loader, not this node, and it means the preset's exact filename isn't in the right ComfyUI/models/ folder. The pack doesn't download model files for you.
• Old SDXL habits. Don't paste a wall of quality tags; the KB has A/B tests showing masterpiece, 8k, ultra detailed is indistinguishable from nothing on these encoders. Describe the scene, keep it tight.