[GH-ISSUE #126] output image resolution details issues

bsakaguchi commented

2026-10-26 21:12:01 +00:00

Owner

Originally created by @bjohn22 on GitHub (Feb 2, 2025).
Original GitHub issue: https://github.com/deepseek-ai/Janus/issues/126

Any thoughts on why I am not able to reproduce the same resolution reported out there:
Weights: I downloaded weights from Huggingface to local and loaded it from local directory
Model: Janus-Pro-7B

Prompt python "/home/mytemp/Documents/ml_projects/Janus/janus_run_script.py" generate --prompt "Elephant on Dallas Texas street" --seed 142 --guidance 5.0

Here is my text to image snippet:

def generate(input_ids,
width,
height,
temperature: float = 1,
parallel_size: int = 16,
cfg_weight: float = 5,
image_token_num_per_image: int = 576,
patch_size: int = 16):
"""
Internal method that generates image patches from the text input.
"""
torch.cuda.empty_cache()
tokens = torch.zeros((parallel_size * 2, len(input_ids)), dtype=torch.int).to(cuda_device)

# Prepare token embeddings
for i in range(parallel_size * 2):
    tokens[i, :] = input_ids
    # Odd rows get "unconditional" tokens
    if i % 2 != 0:
        tokens[i, 1:-1] = vl_chat_processor.pad_id

inputs_embeds = vl_gpt.language_model.get_input_embeddings()(tokens)
generated_tokens = torch.zeros((parallel_size, image_token_num_per_image), dtype=torch.int).to(cuda_device)

pkv = None
# Loop to generate tokens for image
for i in range(image_token_num_per_image):
    outputs = vl_gpt.language_model.model(inputs_embeds=inputs_embeds, use_cache=True, past_key_values=pkv)
    pkv = outputs.past_key_values
    hidden_states = outputs.last_hidden_state
    logits = vl_gpt.gen_head(hidden_states[:, -1, :])

    # Extract conditional/unconditional logits
    logit_cond = logits[0::2, :]
    logit_uncond = logits[1::2, :]
    logits = logit_uncond + cfg_weight * (logit_cond - logit_uncond)

    probs = torch.softmax(logits / temperature, dim=-1)
    next_token = torch.multinomial(probs, num_samples=1)
    generated_tokens[:, i] = next_token.squeeze(dim=-1)

    # Expand the next_token for conditional/unconditional
    next_token = torch.cat([next_token.unsqueeze(dim=1), next_token.unsqueeze(dim=1)], dim=1).view(-1)
    img_embeds = vl_gpt.prepare_gen_img_embeds(next_token)
    inputs_embeds = img_embeds.unsqueeze(dim=1)

# Decode tokens into patches
patches = vl_gpt.gen_vision_model.decode_code(
    generated_tokens.to(dtype=torch.int),
    shape=[parallel_size, 8, width // patch_size, height // patch_size]
)
return generated_tokens.to(dtype=torch.int), patches

def unpack(dec, width, height, parallel_size=16):
"""
Convert raw patches into final numpy images.
"""
dec = dec.to(torch.float32).cpu().numpy().transpose(0, 2, 3, 1)
dec = np.clip((dec + 1) / 2 * 255, 0, 255)

visual_img = np.zeros((parallel_size, width, height, 3), dtype=np.uint8)
visual_img[:, :, :] = dec
return visual_img

@torch.inference_mode()
def generate_image(prompt, seed, guidance):
"""
Generate multiple images for a given prompt.
"""
torch.cuda.empty_cache()
seed = seed if seed is not None else 12345
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
np.random.seed(seed)

width = 384
height = 384
parallel_size = 16

with torch.no_grad():
    # Prepare text
    messages = [
        {'role': 'User', 'content': prompt},
        {'role': 'Assistant', 'content': ''}
    ]
    text = vl_chat_processor.apply_sft_template_for_multi_turn_prompts(
        conversations=messages,
        sft_format=vl_chat_processor.sft_format,
        system_prompt=''
    )
    text = text + vl_chat_processor.image_start_tag

    input_ids = torch.LongTensor(tokenizer.encode(text)).to(cuda_device)

    _, patches = generate(
        input_ids,
        width // 16 * 16,
        height // 16 * 16,
        cfg_weight=guidance,
        parallel_size=parallel_size
    )
    images = unpack(patches, width // 16 * 16, height // 16 * 16)

    # Convert to PIL and upscale
    return [
        Image.fromarray(images[i]).resize((1024, 1024), Image.LANCZOS)
        for i in range(parallel_size)
    ]

Originally created by @bjohn22 on GitHub (Feb 2, 2025). Original GitHub issue: https://github.com/deepseek-ai/Janus/issues/126 Any thoughts on why I am not able to reproduce the same resolution reported out there: Weights: I downloaded weights from Huggingface to local and loaded it from local directory Model: `Janus-Pro-7B` Prompt `python "/home/mytemp/Documents/ml_projects/Janus/janus_run_script.py" generate --prompt "Elephant on Dallas Texas street" --seed 142 --guidance 5.0` Here is my text to image snippet: def generate(input_ids, width, height, temperature: float = 1, parallel_size: int = 16, cfg_weight: float = 5, image_token_num_per_image: int = 576, patch_size: int = 16): """ Internal method that generates image patches from the text input. """ torch.cuda.empty_cache() tokens = torch.zeros((parallel_size * 2, len(input_ids)), dtype=torch.int).to(cuda_device) # Prepare token embeddings for i in range(parallel_size * 2): tokens[i, :] = input_ids # Odd rows get "unconditional" tokens if i % 2 != 0: tokens[i, 1:-1] = vl_chat_processor.pad_id inputs_embeds = vl_gpt.language_model.get_input_embeddings()(tokens) generated_tokens = torch.zeros((parallel_size, image_token_num_per_image), dtype=torch.int).to(cuda_device) pkv = None # Loop to generate tokens for image for i in range(image_token_num_per_image): outputs = vl_gpt.language_model.model(inputs_embeds=inputs_embeds, use_cache=True, past_key_values=pkv) pkv = outputs.past_key_values hidden_states = outputs.last_hidden_state logits = vl_gpt.gen_head(hidden_states[:, -1, :]) # Extract conditional/unconditional logits logit_cond = logits[0::2, :] logit_uncond = logits[1::2, :] logits = logit_uncond + cfg_weight * (logit_cond - logit_uncond) probs = torch.softmax(logits / temperature, dim=-1) next_token = torch.multinomial(probs, num_samples=1) generated_tokens[:, i] = next_token.squeeze(dim=-1) # Expand the next_token for conditional/unconditional next_token = torch.cat([next_token.unsqueeze(dim=1), next_token.unsqueeze(dim=1)], dim=1).view(-1) img_embeds = vl_gpt.prepare_gen_img_embeds(next_token) inputs_embeds = img_embeds.unsqueeze(dim=1) # Decode tokens into patches patches = vl_gpt.gen_vision_model.decode_code( generated_tokens.to(dtype=torch.int), shape=[parallel_size, 8, width // patch_size, height // patch_size] ) return generated_tokens.to(dtype=torch.int), patches def unpack(dec, width, height, parallel_size=16): """ Convert raw patches into final numpy images. """ dec = dec.to(torch.float32).cpu().numpy().transpose(0, 2, 3, 1) dec = np.clip((dec + 1) / 2 * 255, 0, 255) visual_img = np.zeros((parallel_size, width, height, 3), dtype=np.uint8) visual_img[:, :, :] = dec return visual_img @torch.inference_mode() def generate_image(prompt, seed, guidance): """ Generate multiple images for a given prompt. """ torch.cuda.empty_cache() seed = seed if seed is not None else 12345 torch.manual_seed(seed) torch.cuda.manual_seed(seed) np.random.seed(seed) width = 384 height = 384 parallel_size = 16 with torch.no_grad(): # Prepare text messages = [ {'role': 'User', 'content': prompt}, {'role': 'Assistant', 'content': ''} ] text = vl_chat_processor.apply_sft_template_for_multi_turn_prompts( conversations=messages, sft_format=vl_chat_processor.sft_format, system_prompt='' ) text = text + vl_chat_processor.image_start_tag input_ids = torch.LongTensor(tokenizer.encode(text)).to(cuda_device) _, patches = generate( input_ids, width // 16 * 16, height // 16 * 16, cfg_weight=guidance, parallel_size=parallel_size ) images = unpack(patches, width // 16 * 16, height // 16 * 16) # Convert to PIL and upscale return [ Image.fromarray(images[i]).resize((1024, 1024), Image.LANCZOS) for i in range(parallel_size) ] ![Image](https://github.com/user-attachments/assets/6fa2876c-5fc6-4df6-bf66-dde5b8dbdcf5)

bsakaguchi commented

2026-10-26 21:12:02 +00:00

Author

Owner

@bjohn22 commented on GitHub (Feb 2, 2025):

Here are my files list

@bjohn22 commented on GitHub (Feb 2, 2025): ![Image](https://github.com/user-attachments/assets/daa7315c-35a4-4b7d-8922-35da4b87ee35) Here are my files list

bsakaguchi commented

2026-10-26 21:12:03 +00:00

Author

Owner

@ukaprc commented on GitHub (Feb 6, 2025):

Your image is 1024 X 1024 which is larger than the usual 768 X 768. The process uses the Image.LANCZOS method to resize from 384 X 384. Did you change the resize size? Even if you did, this method isn't as effective as using a real upscaler (as they've done in their online demo) such as RealESRGAN. You can easily enough integrate the upscaler into your inference code.

@ukaprc commented on GitHub (Feb 6, 2025): Your image is 1024 X 1024 which is larger than the usual 768 X 768. The process uses the Image.LANCZOS method to resize from 384 X 384. Did you change the resize size? Even if you did, this method isn't as effective as using a real upscaler (as they've done in their online demo) such as RealESRGAN. You can easily enough integrate the upscaler into your inference code.

bsakaguchi commented

2026-10-26 21:12:03 +00:00

Author

Owner

@zohebnsr commented on GitHub (Mar 1, 2025):

The model can't make more than 384 pixel originally? right?

@zohebnsr commented on GitHub (Mar 1, 2025): The model can't make more than 384 pixel originally? right?

bsakaguchi commented

2026-10-26 21:12:03 +00:00

Author

Owner

@ukaprch commented on GitHub (Mar 3, 2025):

zohebnsr
The model is fixed at 384 and then resizes, but you can upscale vs resizing for a better image.

@ukaprch commented on GitHub (Mar 3, 2025): [zohebnsr](https://github.com/zohebnsr) The model is fixed at 384 and then resizes, but you can upscale vs resizing for a better image.

Rows
Columns

[GH-ISSUE #126] output image resolution details issues #95