⚡ TA KSampler

If you've ever wanted to switch your sampler without opening a dropdown on the KSampler, this is the node. TA KSampler is the sampling workhorse of the TA pack, and its one structural difference from ComfyUI's built-in KSampler is the whole point: sampler_name and scheduler are plain text inputs instead of fixed menus. That means you can drive them from another node - most obviously the pack's TA Sampler Preset, but honestly any string source will do.

Everything else is a familiar KSampler. You feed it model, positive and negative conditioning, latent_image, a noise_seed, steps, and cfg, and it hands you a denoised latent out the other end, ready for a VAE Decode. The inputs a beginner actually touches are steps, cfg, and the seed. The fields that make this more than a clone are start_at_step and end_at_step (for stepping through a schedule or hooking into hires workflows), add_noise (set to "disable" for img2img passes where you're already adding noise upstream), and return_with_leftover_noise (handy if you're feeding a refiner that wants to finish the denoise). The preview boolean is on by default and gives you a live look at the latent after every step, right on the node - the same latent2rgb trick the built-in preview uses, no VAE required.

Under the hood it's just ComfyUI's own sampling machinery with a string-typed sampler/scheduler and a per-step progress callback. The preview runs through comfy.utils.ProgressBar and is deliberately fail-silent, so a preview hiccup won't kill a generation. It also tolerates both plain 4-channel latents and the packed 5D ones from video/Qwen models, which is a nice touch if you wander outside plain SD.

The natural setup is the pair the pack documents: TA Sampler Preset feeds steps, cfg, sampler_name, scheduler, and the step bounds as wires; TA KSampler consumes them. Flip the preset and your sampler settings travel with it, no rewiring. A preset like Z-Image Turbo resolves to CFG 1 with a beta scheduler at 9 steps - which is exactly the sort of distilled, flow-matching setting that would trip you up if you guessed. If you're not using the preset, remember these are free-text fields: euler, dpmpp_2m, karras must be spelled exactly or the node errors out.

Install. The pack installs with no extra Python dependencies at all - it ships clean, no requirements.txt. Either grab it from ComfyUI Manager (search "TA ComfyUI Nodes Pack") or:

cd ComfyUI/custom_nodes
git clone https://github.com/tmode-1960/TA-ComfyUI-Nodes-Pack

Restart ComfyUI and the nodes appear under TA Nodes/Sampling.

Gotchas. Because the pack hit a v2.x breaking change, any workflow built against TA v1 nodes needs rebuilding - node names and wiring changed. If you cloned an old workflow and things come up missing, that's why. Also, the sampler/scheduler strings are validated against ComfyUI's lists, so a typo means a hard failure rather than a silently different schedule. And if you see the progress preview disappear mid-run, check the preview toggle - it's the only thing between you and a live peek.