How to replicate Ostris AI toolkit sampler in sdnext?

[SAC020]

Not logging this as a bug because I'm not sure if it's a bug or a configuration issue.

While training loras with AI Toolkit (Qwen 2512 in this case, quantized with an "accuracy recovery adapter") I find myself unable to replicate the training sample outputs in sdnext.

The lora seems to work perfect in the AI Toolkit samples, but it's performing very poorly in sdnext.

To test it more to the root cause, I did a comparison of outputs at zero training steps (pre-training baseline, core model, no lora yet):

AI Toolkit

Prompt: sacbf, candid documentary photo, a woman (sacbf will be the future token, but it's pre-training)

    seed: 4245461
    guidance_scale: 4
    sample_steps: 25
    noise_scheduler: "flowmatch"
    name_or_path: "Qwen/Qwen-Image-2512"
    quantize: true
    quantize_te: true
    arch: "qwen_image:2512"
    low_vram: true
    model_kwargs: {}
    layer_offloading: false
    qtype: "uint4|ostris/accuracy_recovery_adapters/qwen_image_2512_torchao_uint4.safetensors"
    qtype_te: "qfloat8"

The same prompt/seed/guidance scale/steps in sdnext, on-the-fly sdnq quantization int6. Zero resemblance, besides the greenish background.

13:39:53-983284 DEBUG    Sampler: "Euler FlowMatch" class=FlowMatchEulerDiscreteScheduler config={'num_train_timesteps':
                         1000, 'shift': 3, 'use_dynamic_shifting': True, 'base_shift': 0.5, 'max_shift': 1.15,
                         'base_image_seq_len': 256, 'max_image_seq_len': 4096, 'time_shift_type': 'exponential'}
13:39:53-985268 INFO     Processing modifiers: apply
13:39:53-990228 INFO     Base: pipeline=QwenImagePipeline task=TEXT_2_IMAGE batch=1/1x1 set={'prompt': 1,
                         'negative_prompt': 1, 'guidance_scale': 4, 'generator': 'cuda:[4245461]',
                         'num_inference_steps': 25, 'true_cfg_scale': 0, 'output_type': 'np', 'width': 1024, 'height':
                         1024, 'parser': 'fixed'}
13:39:54-000644 DEBUG    Encode: prompt="['sacbf, candid documentary photo, a woman\n']" hijack=True
Progress  4.71s/it ████████████▏                       36% 9/25 00:42 01:15 Base13:40:39-419088 TRACE    Server: alive=True requests=844 memory=46.52/63.92 status='idle' task='' timestamp=None
                         current='task(1k9s528f94rjgez)' id='' job=0 jobs=0 total=52 step=9 steps=0 queued=0 uptime=1226
                         elapsed=43.62 eta=None progress=0
Progress  4.69s/it █████████████████████████████████ 100% 25/25 01:57 00:00 Base
13:41:54-160646 DEBUG    Decode: vae=AutoencoderKLQwenImage slicing=False tiling=False latents=[1, 16, 1, 128,
                         128]:cuda:0 dtype=torch.bfloat16 time=1.231

Any idea how to approach this?

[SAC020]

To simplify more: Z-image turbo, no quantization at all

AI Toolkit

        noise_scheduler: "flowmatch"
          - prompt: "full body view photo of a hiker woman on a mountain trail wearing hiking equipment and boots"
        neg: ""
        seed: 4245456
        walk_seed: false
        guidance_scale: 1
        sample_steps: 8

.

SD-Next - Z-image turbo, no quantization

It's closer, but far from identical. Default sampler behaves identical with EulerFlowmatch in SD-Next

Prompt: full body view photo of a hiker woman on a mountain trail wearing hiking equipment and boots

Parameters: Steps: 8| Size: 1024x1024| Sampler: Euler FlowMatch| Seed: 4245456| CFG scale: 1.0| App: SD.Next| Version: 47dcab3| Pipeline: ZImagePipeline| Operations: txt2img| Model: Z-Image-Turbo| Sampler low order: False| Sampler dynamic shift: True

Time: 11.55s | total 23.21 pipeline 7.72 decode 3.37 onload 3.26 preview 2.99 offload 2.78 te 1.63 callback 0.88 post 0.41 | GPU 14270 MB 44% | RAM 43.29 GB 68%