slop-stuff cheatsheets & experiments git

slop-stuff / Graphics

ComfyUI nodes

Wire the graph — don't pray to the UI.

Node herding for SlopGen ComfyUI — core graph, samplers, loaders, ControlNet, video packs, kill-list.

ComfyUIlatentsamplerFluxLTXV

Cheatsheet for SlopGen ComfyUI (object_info snapshot: 1195 nodes). Core graph first; then what’s actually installed on that box (AnimateDiff, LTXV, Wan, partner APIs, …). Custom packs come and go — treat pack sections as inventory notes, not eternal API.

Quick reference

Minimal txt2img (SD/Flux-style):

CheckpointLoaderSimple → CLIPTextEncode(+) / CLIPTextEncode(−)
                      → EmptyLatentImage
                      → KSampler → VAEDecode → SaveImage
         MODEL/CLIP/VAE ──────────┘
KnobTypicalNotes
steps20–30fewer with distilled/turbo; LCM ~4–8
cfg3.5–7 Flux · ~7–8 SDXL default in UItoo high → fried / overcooked
denoise1.0 txt2img · 0.35–0.75 img2imgstructure from init latent
latent sizemultiple of 8SDXL often 1024²; Flux often 1024+
schedulerkarras / normal / simplepair with sampler; don’t cargo-cult

Type wires that matter: MODEL · CLIP · VAE · CONDITIONING · LATENT · IMAGE · MASK · CONTROL_NET · LORA

Graph anatomy

StageJobCore nodes (this install)
Loadweights into graphCheckpointLoaderSimple, UNETLoader, CLIPLoader, VAELoader, LoraLoader, ControlNetLoader, UpscaleModelLoader
Conditiontext / control → CONDITIONINGCLIPTextEncode, CLIPTextEncodeSDXL, CLIPTextEncodeFlux, ControlNetApplyAdvanced, FluxGuidance
Latentcanvas / encode pixelsEmptyLatentImage, EmptySD3LatentImage, EmptyFlux2LatentImage, VAEEncode, VAEEncodeForInpaint
SampledenoiseKSampler, KSamplerAdvanced, SamplerCustomAdvanced
Decodelatent → pixelsVAEDecode, VAEDecodeTiled
I/Odisk / previewLoadImage, SaveImage, PreviewImage

KEY: Checkpoint loaders spit MODEL + CLIP + VAE. Flux/SD3-style graphs often load UNET/CLIP/VAE separately — don’t assume one .safetensors always means all three.

Loaders

NodeOutputsUse
CheckpointLoaderSimpleMODEL, CLIP, VAEsingle-file ckpt
UNETLoader / CLIPLoader / VAELoaderpiecesFlux / modular pipelines
DualCLIPLoader / TripleCLIPLoaderCLIPmulti-encoder models
LoraLoaderMODEL, CLIPstrength_model / strength_clip (can be negative)
LoraLoaderModelOnlyMODELskip CLIP patch
ControlNetLoaderCONTROL_NETpair with Apply*
UpscaleModelLoaderUPSCALE_MODELwith ImageUpscaleWithModel

Checkpoints seen on SlopGen (examples): flux1-dev-fp8, sd_xl_base_1.0, ltx-2.3-22b-…, hunyuan_3d_v2.1, DynamiCrafter/ToonCrafter. Refresh via the loader dropdown — inventory changes.

Conditioning

NodeRole
CLIPTextEncodegeneric positive/negative text
CLIPTextEncodeSDXL / …Refiner / …SD3architecture-specific text encode
CLIPTextEncodeFlux + FluxGuidanceFlux text + guidance helper
ControlNetApplyAdvancedpos/neg + image + strength + start/end %
InstructPixToPixConditioningedit conditioning
GLIGENTextBoxApplygrounded boxes
PhotoMakerEncodeidentity stack

ControlNet: preprocess the guide image outside or via detection nodes (image/detection), then ControlNetApplyAdvanced — prefer Advanced over legacy ControlNetApply.

Samplers & schedulers

KSampler inputs: model, seed, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, denoise.

Samplers on this build (subset of the long enum):
euler, euler_ancestral, heun, dpm_2, dpmpp_2m, dpmpp_2m_sde, dpmpp_3m_sde, ddim, uni_pc, lcm, res_multistep, … (+ many _cfg_pp / _gpu variants).

Schedulers: simple, sgm_uniform, karras, exponential, ddim_uniform, beta, normal, linear_quadratic, kl_optimal.

GoalTry
General SDXLdpmpp_2m + karras
Ancestral variationeuler_ancestral / dpmpp_2s_ancestral
Few-step / LCM LoRAlcm + low steps
Fine controlKSamplerAdvanced (start/end step, leftover noise)

denoise < 1 needs a meaningful init latent (VAEEncode or previous sample) — otherwise you’re just dimming noise.

Latent & VAE

NodeNotes
EmptyLatentImagew/h step 8; batch_size
VAEEncode / VAEDecodematch VAE to model family
VAEEncodeForInpainthole + mask aware
VAEDecodeTiled / VAEEncodeTiledVRAM lifeline
SetLatentNoiseMaskinpaint / regional
LatentUpscale / LatentUpscaleBybefore second sample pass

Wrong VAE (SD1.5 VAE on SDXL, etc.) → gray sludge / pink nonsense. When in doubt, use the VAE from the same loader stack.

Image I/O & utils

LoadImage · LoadImageMask · SaveImage · PreviewImage · ImageScale / ImageScaleBy · ImageUpscaleWithModel · ImageCompositeMasked · RemoveBackground · mask ops under image/mask.

Workflow recipes

txt2img

Loader → +/− encode → EmptyLatent → KSampler (denoise 1) → VAEDecode → Save

img2img

LoadImage → VAEEncode → KSampler (denoise 0.35–0.75) → Decode

Inpaint

LoadImage + mask → VAEEncodeForInpaint (or encode + SetLatentNoiseMask) → sample → decode

2-pass upscale

Sample → Decode → ImageUpscaleWithModel or latent upscale → second KSampler at lower denoise

ControlNet

Guide image → (optional preprocessor) → ControlNetLoaderControlNetApplyAdvanced on +/− → sample

SlopGen inventory notes

Pulled from live object_info (counts approximate by name prefix/category):

Family~NodesWhat it’s for
AnimateDiff (ADE_*)~143motion modules, context opts, LoRA hooks
LTXV / Lightricks~84+LTX video latents, samplers, AV
Wan*~41Wan video I2V / control / audio dance
ComfyUI-Frame-InterpolationVFIframe interp loaders/nodes
DynamiCrafterWrapperdozenI2V / toon crafter style
partner/manyhosted image/video/3d/audio APIs (Kling, Recraft, Tripo, …)
Woosh/fewcustom load/encode/sample helpers
3d / mesh / splatdozensmeshing & Gaussian-ish paths
audio / QwenTTSpresentTTS / audio encode paths

Partner nodes need API credentials and are not local diffusion — don’t debug them like KSampler graphs.

AI-slop kill-list

SlopReality
Random CFG 15 “for sharpness”Overcooks; use model-appropriate CFG
Any VAE with any UNETFamily must match
denoise=1 on img2img and expecting the photoYou’re reinventing txt2img
Ignoring latent multiple-of-8Subtle crops / errors
Mixing SDXL CLIP encode with Flux UNETGraph type-checks may pass; results won’t
“More steps always better”Plateaus; wastes time
Forgetting negative prompt on SD-familyOften matters; Flux stacks differ
ControlNet strength 2.0 foreverStart ~0.6–1.0; use start/end %
Tiled VAE “always”Slower; use when VRAM hurts
Assuming custom ADE/LTXV nodes are core ComfyPack-specific; break when pack updates
Partner node failures = sampler bugCheck API keys / quotas
Publishing every LoRA filename from the boxInventory is private — don’t mirror NSFW lists into public docs

Refs

  • ComfyUI · examples in-repo
  • Live graph UI: slopgen.jurip.xyz (LAN / HTTP from your network)
  • Node schema: GET /object_info on that host
  • AnimateDiff pack docs (Kosinkadink) · LTXV / Wan model cards for architecture-specific encodes