<p>A bit level image encryption algorithm is proposed in this paper, which utilizes chaotic sequences generated by a denoising diffusion probabilistic model (DDPM) to enhance the security of image data. We choose the Lorenz system to generate the training dataset for the improved DDPM. To generate long and stable chaotic sequences, the model adopts a Transformer encoder-decoder architecture as backbone and is embedded with polynomial and periodic fitting modules. Experimental shows that the chaotic sequences generated by the model outperform the original Lorenz system in both NIST tests and increase the Lyapunov exponent by about 0.6. To further enhance the security of image encryption, we design an advanced logistic chaotic map with a dynamic parameter adjustment mechanism, which achieves better performance than the original system in NIST tests. Additionally, we propose a method to generate random bit sequences from floating-point numbers, and combine this with an improved Feistel cipher system, using the generated chaotic sequences as keys for bit-level image encryption. The effectiveness of the proposed encryption system is evaluated through various experiments, including entropy of 7.9993 and 0.0305 s for encrypting a 256<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation>256 image, indicating its strong capability and performance.</p>