The goal of this pack is to keep genuinely useful glue nodes in one place without pulling in a large third-party custom-node bundle just to get one tiny piece of functionality.

Companion use with LatentSlate

This pack is intended to be a stable companion utility pack for LatentSlate-oriented ComfyUI workflows while remaining useful on its own.

The nodes here are intentionally small, stable, and dependency-light. They cover workflow glue that is useful around provider-driven image generation, image-to-video, resizing, resolution snapping, color matching, local LLM calls, and narrow PyTorch model-setting patches.

This repo does not claim to be a required dependency for LatentSlate. If a LatentSlate workflow package needs one of these nodes, that workflow should list the dependency explicitly. Planned workflow package notes live in workflows/README.md.

Included

`Text Split (Delimiter)`

Splits an input string using a delimiter and returns:

part_1 through part_20
remainder
count

Behavior:

output_count controls how many leading parts are emitted before overflow is collected into remainder
strip_parts trims whitespace around each part
skip_empty removes empty parts after splitting
delimiter supports common escape strings such as \n, \r\n, \r, and \t

This node intentionally uses a fixed socket layout instead of dynamic outputs so saved workflows stay predictable and easy to debug.

`Lazy Index Switch`

Selects one of input_0 through input_19 using a zero-based index value.

Behavior:

accepts any ComfyUI type on its input sockets
evaluates only the selected input branch
pairs with combo-style nodes that expose an INDEX output
reports a clear validation error if the selected input is not connected

`Enviral Krea2 Seed Variance Enhancer`

Adds seeded random noise to CLIP text conditioning embeddings for Krea2, Z-Image Turbo, and similar workflows where prompt-embedding variation can help break repeated compositions.

Inputs:

conditioning: CLIP text conditioning to modify
randomize_percent: percentage of embedding values that receive noise
strength: noise scale; add 1,000,000,000 to use the upstream v2.1 seed behavior
noise_insert: applies noisy conditioning to beginning steps, ending steps, all steps, or disables the node
steps_switchover_percent: sampler-step split point when using beginning or ending step modes
seed: deterministic seed for noise values and mask selection
mask_starts_at and mask_percent: protect part of the prompt embedding from noise
log_to_console: prints tensor statistics and suggested strength ranges

This is adapted from ChangeTheConstants' MIT-0 SeedVarianceEnhancer v2.2.0. The Enviral version avoids mutating PyTorch's global RNG state and guards untagged conditioning metadata so it behaves more predictably inside larger node chains.

Typed Name Selectors

Adapts linked strings, including strings from ComfyUI's built-in Custom Combo node, into exact native combo output types for common model and sampler fields.

Included nodes:

Enviral Sampler Name
Enviral Scheduler Name
Enviral Checkpoint Name
Enviral Diffusion Model Name
Enviral CLIP Name
Enviral CLIP Type
Enviral VAE Name
Enviral LoRA Name
Enviral Color Match Method
Enviral Color Match V2 Method

Behavior:

each node has a normal dropdown for its native ComfyUI option list
each node also accepts a linked string matching one current ComfyUI option exactly, or by an unambiguous case/path-separator/path-suffix match
the output socket uses the exact native combo type expected by downstream nodes such as KSampler, Load Checkpoint, Load Diffusion Model, Load CLIP, Load VAE, and native LoRA loaders
invalid linked strings fail validation before the downstream loader or sampler runs, with close-match suggestions when available

Use Enviral Color Match Method for legacy ColorMatch nodes with CPU methods: mkl, hm, reinhard, mvgd, hm-mvgd-hm, and hm-mkl-hm. Use Enviral Color Match V2 Method for V2 color match nodes that also expose reinhard_lab_gpu.

`Enviral Image Resize Kit`

Resizes an image with a small set of explicit workflow-friendly modes:

pass-through: keeps the input image resolution and ignores width/height
explicit resize: stretches the image to the requested width and height
inside fit: preserves the full image, pads to the requested aspect ratio, fills padding with letterbox_color, then outputs the requested width and height
outside fit: fills the requested aspect ratio and center-crops overflow

Inputs:

image: input image batch
mode: resize mode dropdown or linked string alias
width and height: requested output dimensions, ignored by pass-through
letterbox_color: used only by inside fit; accepts simple names, R,G,B, 0.0-1.0 float triples, or #RRGGBB
optional mask: transformed with the same geometry as the image

The mode input is a dropdown, but it also accepts linked strings such as resize, pad, crop, and pass-through. This makes it pair cleanly with combo nodes that output strings. Optional masks are transformed with the same geometry and the node returns image, width, height, and mask.

`Enviral Color Match V2`

Transfers color from a reference image batch to a target image batch.

Inputs:

image_target: image batch to recolor
image_ref: image batch to sample color from
method: mkl, hm, reinhard, mvgd, hm-mvgd-hm, hm-mkl-hm, or reinhard_lab_gpu
strength: blend/extrapolation amount, where 0 is unchanged and 1 is the matched result
multithread: enables threaded CPU processing for multi-image batches

The CPU methods use color-matcher. The reinhard_lab_gpu method uses ComfyUI's Kornia dependency for Lab-space matching and requires both inputs to be RGB IMAGE tensors with 3 channels. The method input can be driven by Enviral Color Match V2 Method when the choice needs to come from a string-producing node.

`Enviral Load LoRA`

LoRA loader with folder and text allow-list filtering, up to five independently bypassable LoRA banks, and a text output containing the final visible list. The folder and LoRA choices come from ComfyUI's registered LoRA paths, including paths supplied through extra_model_paths.yaml.

Inputs:

model: diffusion model to patch
clip: optional CLIP model to patch; leave disconnected for model-only LoRAs
folder: All LoRAs or any current LoRA folder prefix; nested folders are included in their parent filter
banks: number of visible LoRA banks, from one to five
allow_list: optional STRING input intended for ComfyUI's native Text (Multiline) node; each non-empty line keeps an exact displayed name or full relative LoRA path
each bank has a LoRA dropdown plus separate model and CLIP strengths; a bank is bypassed when both strengths are zero; changing banks adds or removes complete bank controls while retaining their values during the session

Outputs:

patched model
patched clip
lora_list: newline-separated final dropdown options; names are relative to the selected folder, or include full relative paths under All LoRAs

The node uses ComfyUI's normal combo refresh behavior. Press R on the canvas or use Refresh Node Definitions to re-enumerate LoRAs and folders, then reapply the folder and allow-list filters. Saved selections are preserved when still valid and otherwise move to the first visible option. Refreshes do not resize the node; only changing the number of visible banks adjusts its height.

The earlier standard and model-only Enviral LoRA loaders remain registered for existing workflows but are deprecated and hidden from normal node discovery.

Filtered Model Loaders

Three single-selection loaders use the same folder, allow-list, relative-label, and visible-list behavior as Enviral Load LoRA:

Enviral Load Checkpoint returns model, clip, vae, and checkpoint_list.
Enviral Load Diffusion Model returns model and diffusion_model_list, while retaining the native weight_dtype choices.
Enviral Load Text Encoder returns clip and text_encoder_list, while retaining the native encoder type and optional device choices.

Each has one selection only—there are no banks. The selection continues to use the full relative ComfyUI path for loading; when a folder is chosen, the dropdown and list output omit that redundant folder prefix.

`Model Patch Torch Settings`

Experimental node that patches a compatible ComfyUI MODEL with callbacks for one CUDA PyTorch backend setting.

Inputs:

model: model to clone and patch
enable_fp16_accumulation: toggles torch.backends.cuda.matmul.allow_fp16_accumulation while the model runs

When enabled, the patched model turns full FP16 accumulation on before model execution and turns it back off during cleanup. When disabled, it forces the setting off before model execution. This requires a PyTorch build that exposes torch.backends.cuda.matmul.allow_fp16_accumulation and a ComfyUI model patcher with clone() and add_callback() support.

`WAN Resolution Snap`

Converts a preset or custom typed resolution into divisor-safe dimensions for WAN video workflows.

Inputs:

preset: custom or common square, landscape, portrait, and 4:3 presets
width and height: freeform custom dimensions, used when preset is custom
snap: nearest, floor, or ceil
divisible_by: output divisor, defaulting to 32
swap: swaps width and height before snapping

Outputs:

width and height as integers
width_float and height_float for nodes that expect float sockets
summary, such as 768 x 1344 | 1.03 MP | 4:7 | div32

If snapping changes the requested dimensions, summary also appends the source size, for example | from 1000 x 540.

`LM Studio Unified (URL + API Key)`

Sends text, an image, or both to an LM Studio OpenAI-compatible chat endpoint.

Inputs:

base_url: accepts http://127.0.0.1:1234/v1, https://lmstudio.example.com, or a full /v1/chat/completions URL
api_key: optional bearer token for LM Studio API-token auth
api_key_env_var: optional environment variable fallback when api_key is blank, defaulting to LMSTUDIO_API_KEY. On Windows, this checks the current process first, then the User and Machine environment variable registry values.
model, prompt, system_prompt, seed, max_tokens, temperature, and timeout_seconds
optional image input for vision-capable models. If the input is an IMAGE batch, each frame is sent as a separate image_url content part, capped by max_images
user_agent: HTTP User-Agent override for routes that block Python's default urllib signature
debug: prints request details to the ComfyUI console

For a secured Cloudflare route, use a public base_url such as:

https://lmstudio.enviral-design.com

The node will call:

https://lmstudio.enviral-design.com/v1/chat/completions

Install

Clone this repo into ComfyUI/custom_nodes and restart ComfyUI.

ComfyUI/custom_nodes/comfyUI-enviral-design-node-pack

Most nodes use ComfyUI's existing Python, PyTorch, Pillow, and NumPy environment. Enviral Color Match V2 also uses color-matcher for its CPU methods; that dependency is declared in pyproject.toml. If you install this repo manually and see a missing color_matcher import, install color-matcher into the same Python environment that runs ComfyUI.

Workflow examples

This repo currently documents nodes and planned workflow packages. It does not currently bundle executable workflow JSON files. See workflows/README.md for the planned package list and status.

Donations & Support

If this saves you time, you can support the work here: