Functions — github.com/black-forest-labs/flux2

hub / github.com/black-forest-labs/flux2 / functions

Functions110 in github.com/black-forest-labs/flux2

⨍ Functions 110◇ Types & classes 27

↓ 10 callersMethod__init__

(self, params: Flux2Params)

src/flux2/model.py:53

↓ 7 callersFunctiontimestep_embedding

Create sinusoidal timestep embeddings. :param t: a 1-D Tensor of N indices, one per batch element. These may be fractio

src/flux2/model.py:710

↓ 6 callersMethod__init__

( self, resolution: int, in_channels: int, ch: int, ch_mult: list[int]

src/flux2/autoencoder.py:109

↓ 5 callersFunctionload_qwen3_embedder

(variant: str, device: str | torch.device = "cuda")

src/flux2/text_encoder.py:435

↓ 4 callersFunctionapply_rope

(xq: Tensor, xk: Tensor, freqs_cis: Tensor)

src/flux2/model.py:828

↓ 4 callersFunctioncausal_attn_fn

Causal attention where reference tokens only attend to themselves. Without cache: layout is [txt, ref, img]. txt+img attend to all, ref self

src/flux2/model.py:758

↓ 4 callersMethodcopy

(self)

scripts/cli.py:43

↓ 4 callersFunctionprint_config

(cfg: Config)

scripts/cli.py:148

↓ 4 callersFunctionswish

(x: Tensor)

src/flux2/autoencoder.py:20

↓ 3 callersMethodencode

(self, x: Tensor)

src/flux2/autoencoder.py:314

↓ 3 callersMethodforward_kv_extract

First denoising step with reference tokens. Runs full forward pass and extracts KV cache for reference tokens to reuse on subsequent

src/flux2/model.py:170

↓ 2 callersMethod_apply_residuals

(self, img, txt, img_attn, txt_attn, mods)

src/flux2/model.py:614

↓ 2 callersFunction_blend_mod_triple

Blend a (shift, scale, gate) triple: first num_ref positions get ref_m, rest get img_m.

src/flux2/model.py:329

↓ 2 callersMethod_out

(self, x: Tensor, attn: Tensor, mlp: Tensor, mod_gate: Tensor)

src/flux2/model.py:482

↓ 2 callersMethod_prepare_qkv

(self, img, txt, pe, pe_ctx, mod_img, mod_txt)

src/flux2/model.py:569

↓ 2 callersMethod_qkv

(self, x: Tensor, mod: tuple[Tensor, Tensor, Tensor])

src/flux2/model.py:468

↓ 2 callersFunctionapply_updates

(cfg: Config, updates: Dict[str, Any])

scripts/cli.py:110

↓ 2 callersFunctionbatched_wrapper

(fn)

src/flux2/sampling.py:106

↓ 2 callersFunctioncap_pixels

(img: Image.Image | list[Image.Image], k)

src/flux2/sampling.py:178

↓ 2 callersMethoddecode

(self, z: Tensor)

src/flux2/autoencoder.py:327

↓ 2 callersMethodformat_input

Format a batch of text prompts into the conversation format expected by apply_chat_template. Optionally, add images to the input.

src/flux2/text_encoder.py:75

↓ 2 callersFunctionload_text_encoder

(model_name: str, device: str | torch.device = "cuda")

src/flux2/util.py:129

↓ 2 callersFunctionparse_key_values

Parse shell-like 'key=value' pairs. Values can be quoted. Example: prompt="a dog" width=768 input_images="in1.png,in2.jpg"

scripts/cli.py:127

↓ 2 callersMethodtest_image

(self, image)

src/flux2/text_encoder.py:424

↓ 2 callersFunctionvalidate_model_params

Validate that config parameters match model requirements. Returns True if valid.

scripts/cli.py:212

↓ 1 callersMethod__init__

( self, model_spec: str, device: str | torch.device = "cuda", )

src/flux2/text_encoder.py:367

↓ 1 callersFunction_blend_double_mods

Blend double block modulations (mod1, mod2) for [ref, img] layout.

src/flux2/model.py:346

↓ 1 callersFunction_blend_single_mods

Blend single block modulations for [txt, ref, img] layout.

src/flux2/model.py:356

↓ 1 callersMethod_format_messages

( self, prompt: str, system_message: str, images: list[Image.Image] | None = N

src/flux2/openrouter_api_client.py:39

↓ 1 callersMethod_validate_and_process_images

( self, img: list[list[Image.Image]] | list[Image.Image] )

src/flux2/text_encoder.py:56

↓ 1 callersMethodattention

(self, h_: Tensor)

src/flux2/autoencoder.py:36

↓ 1 callersFunctioncap_min_pixels

(img: Image.Image | list[Image.Image], max_ar=8, min_sidelength=64)

src/flux2/sampling.py:195

↓ 1 callersFunctioncenter_crop_to_multiple_of_x

( img: Image.Image | list[Image.Image], x: int )

src/flux2/sampling.py:159

↓ 1 callersFunctioncoerce_value

Convert a raw string to the correct field type.

scripts/cli.py:67

↓ 1 callersFunctioncompress_time

(t_ids: Tensor)

src/flux2/sampling.py:12

↓ 1 callersFunctioncompute_empirical_mu

(image_seq_len: int, num_steps: int)

src/flux2/sampling.py:251

↓ 1 callersFunctionconcatenate_images

Concatenate a list of PIL images horizontally with center alignment and white background.

src/flux2/sampling.py:413

↓ 1 callersFunctiondefault_images_prep

( x: Image.Image | list[Image.Image], )

src/flux2/sampling.py:217

↓ 1 callersFunctiondefault_prep

( img: Image.Image | list[Image.Image], limit_pixels: int | None, ensure_multiple: int = 16 )

src/flux2/sampling.py:226

↓ 1 callersFunctiondenoise_cfg

( model: Flux2, img: Tensor, img_ids: Tensor, txt: Tensor, # Already cat([txt_empty, txt_prom

src/flux2/sampling.py:364

↓ 1 callersFunctionencode_image_refs

(ae, img_ctx: list[Image.Image])

src/flux2/sampling.py:52

↓ 1 callersMethodforward_kv_cached

Subsequent denoising steps using cached KV for reference tokens. Input x has layout [img] only (no ref tokens).

src/flux2/model.py:267

↓ 1 callersFunctiongeneralized_time_snr_shift

(t: Tensor, mu: float, sigma: float)

src/flux2/sampling.py:240

↓ 1 callersFunctionget_schedule

(num_steps: int, image_seq_len: int)

src/flux2/sampling.py:244

↓ 1 callersFunctionimage_to_base64

Convert PIL Image to base64 string.

src/flux2/util.py:169

↓ 1 callersMethodinv_normalize

(self, z)

src/flux2/autoencoder.py:308

↓ 1 callersFunctionlisted_wrapper

(fn)

src/flux2/sampling.py:122

↓ 1 callersFunctionload_ae

(model_name: str, device: str | torch.device = "cuda")

src/flux2/util.py:134

↓ 1 callersFunctionload_flow_model

(model_name: str, debug_mode: bool = False, device: str | torch.device = "cuda")

src/flux2/util.py:91

↓ 1 callersMethodnormalize

(self, z)

src/flux2/autoencoder.py:304

↓ 1 callersFunctionprint_help

()

scripts/cli.py:168

↓ 1 callersFunctionrope

(pos: Tensor, dim: int, theta: int)

src/flux2/model.py:818

↓ 1 callersFunctionscatter_ids

using position ids to scatter tokens into place

src/flux2/sampling.py:24

↓ 1 callersMethodtest_txt

(self, txt: str)

src/flux2/text_encoder.py:421

↓ 1 callersFunctionto_rgb

(img: Image.Image | list[Image.Image])

src/flux2/sampling.py:206

↓ 1 callersMethodupsample_prompt

Upsample prompts using OpenRouter API. Args: txt: List of input prompts to upsample img: Optional list of im

src/flux2/openrouter_api_client.py:69

↓ 1 callersMethodupsample_prompt

(self, txt: list[str], img=None, **kwargs)

src/flux2/text_encoder.py:427

Method__call__

Adds a predefined watermark to the input image Args: image: ([N,] B, RGB, H, W) in range [-1, 1] Returns:

src/flux2/watermark.py:13

Method__init__

Initialize the OpenRouter API client. Args: model: Model name to use for upsampling. Defaults to "mistralai/pixtral-larg

src/flux2/openrouter_api_client.py:18

Method__init__

(self, watermark)

src/flux2/watermark.py:7

Method__init__

(self, in_channels: int)

src/flux2/autoencoder.py:25

Method__init__

(self, in_channels: int, out_channels: int)

src/flux2/autoencoder.py:55

Method__init__

(self, in_channels: int)

src/flux2/autoencoder.py:85

Method__init__

(self, in_channels: int)

src/flux2/autoencoder.py:98

Method__init__

( self, ch: int, out_ch: int, ch_mult: list[int], num_res_blocks: int,

src/flux2/autoencoder.py:185

Method__init__

(self, params: AutoEncoderParams)

src/flux2/autoencoder.py:272

Method__init__

( self, dim: int, num_heads: int = 8, )

src/flux2/model.py:376

Method__init__

(self)

src/flux2/model.py:391

Method__init__

(self, dim: int, double: bool, disable_bias: bool = False)

src/flux2/model.py:401

Method__init__

( self, hidden_size: int, out_channels: int, )

src/flux2/model.py:416

Method__init__

( self, hidden_size: int, num_heads: int, mlp_ratio: float = 4.0, )

src/flux2/model.py:438

Method__init__

( self, hidden_size: int, num_heads: int, mlp_ratio: float, )

src/flux2/model.py:525

Method__init__

(self, in_dim: int, hidden_dim: int, disable_bias: bool = False)

src/flux2/model.py:684

Method__init__

(self, dim: int, theta: int, axes_dim: list[int])

src/flux2/model.py:695

Method__init__

(self, dim: int)

src/flux2/model.py:735

Method__init__

(self, dim: int)

src/flux2/model.py:747

Method__init__

( self, model_spec: str = "mistralai/Mistral-Small-3.2-24B-Instruct-2506", model_spec_

src/flux2/text_encoder.py:34

Functionbatched_prc

(x: Tensor, t_coord: Tensor | None = None)

src/flux2/sampling.py:107

Functiondenoise

( model: Flux2, # model input img: Tensor, img_ids: Tensor, txt: Tensor, txt_ids: Tens

src/flux2/sampling.py:269

Functiondenoise_cached

Denoise with KV caching for reference image tokens. Step 0: model.forward_kv_extract() — full pass with ref tokens, extracts KV cache. Steps

src/flux2/sampling.py:310

Methodforward

(self, x: Tensor)

src/flux2/autoencoder.py:50

Methodforward

(self, x)

src/flux2/autoencoder.py:68

Methodforward

(self, x: Tensor)

src/flux2/autoencoder.py:90

Methodforward

(self, x: Tensor)

src/flux2/autoencoder.py:102

Methodforward

(self, x: Tensor)

src/flux2/autoencoder.py:159

Methodforward

(self, z: Tensor)

src/flux2/autoencoder.py:239

Methodforward

( self, x: Tensor, x_ids: Tensor, timesteps: Tensor, ctx: Tensor,

src/flux2/model.py:115

Methodforward

(self, x: Tensor)

src/flux2/model.py:395

Methodforward

(self, vec: torch.Tensor)

src/flux2/model.py:407

Methodforward

(self, x: torch.Tensor, vec: torch.Tensor)

src/flux2/model.py:426

Methodforward

(self, x: Tensor)

src/flux2/model.py:690

Methodforward

(self, ids: Tensor)

src/flux2/model.py:701

Methodforward

(self, x: Tensor)

src/flux2/model.py:739

Methodforward

(self, q: Tensor, k: Tensor, v: Tensor)

src/flux2/model.py:752

Methodforward

(self, txt: list[str])

src/flux2/text_encoder.py:218

Methodforward

(self, txt: list[str])

src/flux2/text_encoder.py:384

Methodforward_kv_cached

Forward using cached ref KV. Input x has layout [txt, img] (no ref).

src/flux2/model.py:508

Methodforward_kv_cached

Forward using cached ref KV. img has layout [img] only (no ref).

src/flux2/model.py:663

Methodforward_kv_extract

Forward with causal attention. Extracts and returns ref KV cache.

src/flux2/model.py:486

Methodforward_kv_extract

Forward with causal attention. img has layout [ref, img]. Extracts ref KV cache.

src/flux2/model.py:637

next →1–100 of 110, ranked by callers