Nodes/comfyui-anima-slider-node/Train Anima Slider LoRA
ComfyUI Node

Train Anima Slider LoRA

Train an experimental Anima/Cosmos flow-slider LoRA from a prompt YAML using the connected MODEL and CLIP.

By Shiba-2-shiba·Created 3 months ago·Updated 2 months ago· 2
Train Anima Slider LoRA
  • model
  • clip
  • vae
  • lora
  • report_json
  • lora_path
  • report_path
prompt_yamlprompts-anima-breast_size_slider.yaml
custom_prompt_yaml_path
prompt_indices0,1,2,3,4,5
eval_prompt_indices6,7
steps600
lr0
rank16
alpha16.0
network_presetattn_mlp
network_reg_dims
network_reg_lrs
model_residencyprefer_cuda
lora_weight_dtypefp32
gradient_checkpointingtrue
skip_initial_evaltrue
skip_final_evalfalse
width0
height0
num_inference_steps20
timestep_samplingshift
sigmoid_scale1.00
discrete_flow_shift3.00
loss_weighting_schemenone
direction_lossenhance_only
teacher_guidance_scale1.00
teacher_norm_referencepositive
min_step_index-1
max_step_index-1
eval_step_indices
eta1.00
seed961218314523996
eval_seed961218314523996
vary_seedtrue
allow_unsafe_age_termsfalse
output_lora_prefixloras/anima_slider
Categorytraining/anima slider

Inputs (38)

NameTypeDefaultDescription
modelMODELLoaded diffusion model to train against.
clipCLIPLoaded text encoder used to encode the prompt YAML.
vaeVAEAccepted for workflow parity with Anima pipelines; the current text-only trainer does not encode images.
prompt_yamlCOMBOprompts-anima-breast_size_slider.yamlPrompt YAML bundled with this custom node.
custom_prompt_yaml_pathSTRINGOptional absolute or ComfyUI-working-directory-relative YAML path. Overrides prompt_yaml when set.
prompt_indicesSTRING0,1,2,3,4,5Comma-separated prompt indices to cycle during training.
eval_prompt_indicesSTRING6,7Comma-separated prompt indices for before/after eval. Empty uses prompt_indices.
stepsINT6001–100000Training optimizer steps.
lrFLOAT00–1Fallback LoRA learning rate.
rankINT161–256Fallback LoRA rank.
alphaFLOAT16.00–1024LoRA alpha.
network_presetCOMBOattn_mlpLoRA target preset.
network_reg_dimsSTRINGOptional YAML mapping of regex fullmatch patterns to LoRA ranks.
network_reg_lrsSTRINGOptional YAML mapping of regex fullmatch patterns to learning rates.
model_residencyCOMBOprefer_cudaBest-effort base model residency after ComfyUI loading. Falls back to DynamicVRAM behavior if CUDA promotion fails.
lora_weight_dtypeCOMBOfp32Trainable LoRA weight dtype. fp32 is recommended; base/bf16 reduce VRAM but may lose small updates.
gradient_checkpointingBOOLEANtrueCheckpoint trainable diffusion blocks during LoRA training to reduce activation VRAM.
skip_initial_evalBOOLEANtrueSkip the pre-training eval pass for OOM isolation. Not a quality substitute.
skip_final_evalBOOLEANfalseSkip the post-training eval pass for OOM isolation. Not a quality substitute.
widthINT00–4096Training latent width in pixels. 0 uses the selected prompt YAML resolution.
heightINT00–4096Training latent height in pixels. 0 uses the selected prompt YAML resolution.
num_inference_stepsINT203–200Number of simple scheduler sigmas.
timestep_samplingCOMBOshift6 options: uniform, mid, early_late, sigmoid, shift, flux_shift
sigmoid_scaleFLOAT1.000.01–20
discrete_flow_shiftFLOAT3.000.01–20
loss_weighting_schemeCOMBOnone3 options: none, sigma_sqrt, cosmap
direction_lossCOMBOenhance_only2 options: enhance_only, bidirectional
teacher_guidance_scaleFLOAT1.000–20Global multiplier applied after each prompt YAML guidance_scale.
teacher_norm_referenceCOMBOpositiveOutput norm reference for the teacher signal. neutral usually makes stronger sliders less prone to scale blow-up.
min_step_indexINT-1-1–10000-1 uses the default lower bound.
max_step_indexINT-1-1–10000-1 uses the default upper bound.
eval_step_indicesSTRINGComma-separated eval step indices. Empty uses midpoint.
etaFLOAT1.000–20
seedINT9612183145239960–18446744073709550000
eval_seedINT9612183145239960–18446744073709550000
vary_seedBOOLEANtrue
allow_unsafe_age_termsBOOLEANfalse
output_lora_prefixSTRINGloras/anima_sliderOutput prefix under the ComfyUI output directory.

Outputs (4)

NameTypeDescription
loraLORA_MODEL
report_jsonSTRING
lora_pathSTRING
report_pathSTRING